https://cognaptus.com/tags/ai-safety/ Recent content in AI Safety on Cognaptus Hugo -- 0.145.0 en-us Mon, 07 Sep 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-09-07-when-the-test-becomes-a-signal-rethinking-ai-agent-evaluation/ Mon, 07 Sep 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-09-07-when-the-test-becomes-a-signal-rethinking-ai-agent-evaluation/ Agent evaluations become less informative when the system can infer that it is being tested and condition its behavior on that inference. https://cognaptus.com/blog/2026-08-14-a-sunset-clause-is-not-a-safety-test-designing-an-exit-from-frontier-ai-limits/ Fri, 14 Aug 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-08-14-a-sunset-clause-is-not-a-safety-test-designing-an-exit-from-frontier-ai-limits/ A frontier-AI treaty needs more than an expiry date: its exit rules must balance genuine risk reduction against verification, sovereignty, and geopolitical change. https://cognaptus.com/blog/2026-08-12-from-alarm-signal-to-release-gate-measuring-cbrn-uplift-in-frontier-models/ Wed, 12 Aug 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-08-12-from-alarm-signal-to-release-gate-measuring-cbrn-uplift-in-frontier-models/ A large controlled CBRN study shows why expert-like harmful output should trigger investigation, not automatically determine a model-release decision. https://cognaptus.com/blog/2026-08-10-shared-blind-spots-why-narrow-ai-needs-an-independent-detector/ Mon, 10 Aug 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-08-10-shared-blind-spots-why-narrow-ai-needs-an-independent-detector/ A formal model shows when narrow systems with escalation create value—and when a monitor sharing the same blind spot makes the safety case collapse. https://cognaptus.com/blog/2026-08-01-the-answer-looked-clinical-the-critical-steps-were-missing/ Sat, 01 Aug 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-08-01-the-answer-looked-clinical-the-critical-steps-were-missing/ A clinician-built evaluation shows why polished medical AI responses can pass presentation checks while omitting the reasoning steps most closely tied to patient harm. https://cognaptus.com/blog/2026-07-30-a-full-distribution-is-not-a-risk-certificate/ Thu, 30 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-30-a-full-distribution-is-not-a-risk-certificate/ A state-level audit shows why distributional reinforcement-learning heads should not drive safety decisions until their strongest risk claims are independently validated. https://cognaptus.com/blog/2026-07-24-agree-once-remember-later-the-commit-boundary-in-personal-agents/ Fri, 24 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-24-agree-once-remember-later-the-commit-boundary-in-personal-agents/ A benchmark of stateful personal agents shows that the largest sycophancy risk emerges when user claims are written into durable state and reused later. https://cognaptus.com/blog/2026-07-22-measure-for-measure-why-ai-evaluation-must-follow-the-failure/ Wed, 22 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-22-measure-for-measure-why-ai-evaluation-must-follow-the-failure/ Two very different AI papers reveal why credible evaluation must trace each intervention from the mechanism it changes to the operational outcome it is meant to improve. https://cognaptus.com/blog/2026-07-16-the-refusal-rate-that-refuses-to-reassure/ Thu, 16 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-16-the-refusal-rate-that-refuses-to-reassure/ A large automated red-team study shows why high aggregate refusal rates can conceal concentrated, inexpensive, and operationally significant jailbreak exposure. https://cognaptus.com/blog/2026-07-09-the-simulator-gets-a-reality-check/ Thu, 09 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-09-the-simulator-gets-a-reality-check/ RealityBridge shows how editable 3DGS driving simulations can become more realistic without letting generative video models rewrite the safety-critical scene. https://cognaptus.com/blog/2026-07-08-the-bike-learns-to-lean-before-it-learns-to-race/ Wed, 08 Jul 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-07-08-the-bike-learns-to-lean-before-it-learns-to-race/ A mechanism-first reading of a self-paced reinforcement-learning framework for autonomous superbike racing, and what it teaches operators about curriculum design in high-dynamics simulation. https://cognaptus.com/blog/2026-06-24-the-big-red-button-is-not-a-risk-model/ Wed, 24 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-24-the-big-red-button-is-not-a-risk-model/ A mechanism-first reading of why shutdown resistance should be diagnosed before enterprises pay for costly, capability-reducing safety designs. https://cognaptus.com/blog/2026-06-17-context-collapse-why-ais-next-bottleneck-is-knowing-what-matters/ Wed, 17 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-17-context-collapse-why-ais-next-bottleneck-is-knowing-what-matters/ A practical synthesis of three arXiv papers showing why fine-grained contextual control, not generic model fluency, is becoming the deployment bottleneck for AI. https://cognaptus.com/blog/2026-06-16-mind-the-tail-quantum-rareevent-sampling-without-the-discovery-tax/ Tue, 16 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-16-mind-the-tail-quantum-rareevent-sampling-without-the-discovery-tax/ A mechanism-first reading of how quantum rare-event sampling changes the cost of finding tail scenarios, and why the advantage depends on aggregate rare-tail mass rather than black-swan theatre. https://cognaptus.com/blog/2026-06-15-cheap-seats-sharp-eyes-rewardhack-detection-without-the-frontier-judge/ Mon, 15 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-15-cheap-seats-sharp-eyes-rewardhack-detection-without-the-frontier-judge/ A small trajectory encoder nearly matches a frontier LLM judge on reward-hack detection, but only when it can read the reasoning-rich trace. https://cognaptus.com/blog/2026-06-12-lie-detectors-are-late-why-ai-oversight-needs-commitment-tracing/ Fri, 12 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-12-lie-detectors-are-late-why-ai-oversight-needs-commitment-tracing/ A mechanism-first reading of counterfactual localization, a method for finding when model reasoning shifts toward deception before the final answer exists. https://cognaptus.com/blog/2026-06-10-trust-issues-benchmarked-why-hallucination-detection-is-a-portfolio-problem/ Wed, 10 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-10-trust-issues-benchmarked-why-hallucination-detection-is-a-portfolio-problem/ OpenHalDet shows why hallucination guardrails should be selected by scenario, model access, and evidence cost—not by a single leaderboard score. https://cognaptus.com/blog/2026-06-08-the-policy-has-to-work-somewhere-rl-for-scale-trust-and-other-inconveniences/ Mon, 08 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-08-the-policy-has-to-work-somewhere-rl-for-scale-trust-and-other-inconveniences/ A business-focused reading of how reinforcement learning can address the two deployment problems that benchmarks politely ignore: distributed scale and trustworthy agent behavior. https://cognaptus.com/blog/2026-06-06-mind-the-slot-jailbreak-prompts-have-weak-points-not-just-bad-words/ Sat, 06 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-06-mind-the-slot-jailbreak-prompts-have-weak-points-not-just-bad-words/ SlotGCG shows that LLM jailbreak risk is shaped not only by adversarial token content, but by where those tokens touch the prompt. https://cognaptus.com/blog/2026-05-30-dont-just-guard-the-door-jailbreak-safety-needs-checkpoints/ Sat, 30 May 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-05-30-dont-just-guard-the-door-jailbreak-safety-needs-checkpoints/ A practical synthesis of three jailbreak-defense papers showing why AI safety should test the path from prompt to response, not just the prompt itself. https://cognaptus.com/blog/2026-05-29-the-confidence-trick-when-long-ai-reasoning-arrives-too-early/ Fri, 29 May 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-05-29-the-confidence-trick-when-long-ai-reasoning-arrives-too-early/ A mechanism-first reading of premature confidence: why longer reasoning traces can still be post-hoc decoration, and how confidence trajectories may help diagnose and train better LLM reasoning. https://cognaptus.com/blog/2026-05-16-context-is-the-new-attack-surface/ Sat, 16 May 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-05-16-context-is-the-new-attack-surface/ A business-focused reading of Jailbreak Mimicry, explaining why LLM safety failures often live in task framing rather than forbidden words. https://cognaptus.com/blog/2026-05-02-jailbreak-at-the-substation-when-grid-ai-learns-the-wrong-shortcut/ Sat, 02 May 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-05-02-jailbreak-at-the-substation-when-grid-ai-learns-the-wrong-shortcut/ A practical reading of a new smart-grid LLM security benchmark, and what it tells business leaders about deploying AI in regulated operations. https://cognaptus.com/blog/2026-04-26-drift-happens-stresstesting-ai-policies-before-sensors-lie/ Sun, 26 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-26-drift-happens-stresstesting-ai-policies-before-sensors-lie/ A practical reading of recent research on measuring how much observation drift an AI policy can tolerate before deployment performance breaks. https://cognaptus.com/blog/2026-04-23-sirens-in-the-weights-why-ai-safety-may-be-hiding-inside-the-model/ Thu, 23 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-23-sirens-in-the-weights-why-ai-safety-may-be-hiding-inside-the-model/ SIREN suggests that harmfulness detection may work better when it listens to internal model representations rather than waiting for a guard model to generate a final label. https://cognaptus.com/blog/2026-04-22-silent-errors-loud-consequences-asmrbench-and-the-coming-era-of-ai-auditors/ Wed, 22 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-22-silent-errors-loud-consequences-asmrbench-and-the-coming-era-of-ai-auditors/ A research-sabotage benchmark shows why AI auditability is not a code-review feature, but an operating model for trustworthy AI work. https://cognaptus.com/blog/2026-04-16-grid-guardians-why-ai-needs-a-safety-chaperone-before-running-the-power-grid/ Thu, 16 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-16-grid-guardians-why-ai-needs-a-safety-chaperone-before-running-the-power-grid/ A mechanism-first reading of why reinforcement learning for power-grid control needs runtime safety shielding, not just better reward penalties. https://cognaptus.com/blog/2026-04-15-benchmarking-the-benchmarks-when-ai-safety-metrics-stop-meaning-anything/ Wed, 15 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-15-benchmarking-the-benchmarks-when-ai-safety-metrics-stop-meaning-anything/ A sharper reading of AISafetyBenchExplorer, showing why AI safety evaluation now suffers less from benchmark scarcity than from metric drift, stale infrastructure, and weak benchmark governance. https://cognaptus.com/blog/2026-04-14-meerkat-or-mirage-when-ai-safety-fails-in-plain-sight-across-traces/ Tue, 14 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-14-meerkat-or-mirage-when-ai-safety-fails-in-plain-sight-across-traces/ A case-first reading of Meerkat shows why AI agent safety failures increasingly require repository-level investigation, not one-trace-at-a-time monitoring. https://cognaptus.com/blog/2026-04-14-when-ai-drives-whos-in-control-reclaiming-determinism-in-agentic-systems/ Tue, 14 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-14-when-ai-drives-whos-in-control-reclaiming-determinism-in-agentic-systems/ A mechanism-first reading of how reactor-based orchestration can make agentic AI safer by bounding nondeterminism instead of pretending to remove it. https://cognaptus.com/blog/2026-04-11-the-cost-of-playing-it-safe-when-ai-safety-creates-harm/ Sat, 11 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-11-the-cost-of-playing-it-safe-when-ai-safety-creates-harm/ A mechanism-first reading of IatroBench, showing how AI safety systems can reduce dangerous outputs while increasing high-stakes omission risk. https://cognaptus.com/blog/2026-04-10-disagreement-is-data-why-ai-needs-more-arguments-not-fewer/ Fri, 10 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-10-disagreement-is-data-why-ai-needs-more-arguments-not-fewer/ A mechanism-first reading of DiADEM shows why subjective AI systems need to model who disagrees, not merely average labels into a convenient fiction. https://cognaptus.com/blog/2026-04-10-when-your-ai-knows-too-little-the-hidden-bottleneck-in-personal-agents/ Fri, 10 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-10-when-your-ai-knows-too-little-the-hidden-bottleneck-in-personal-agents/ KnowU-Bench shows why the next bottleneck for mobile AI agents is not clicking the right button, but acquiring preferences, composing constraints, and knowing when not to intervene. https://cognaptus.com/blog/2026-04-07-the-cost-of-convenience-when-ai-help-becomes-cognitive-debt/ Tue, 07 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-07-the-cost-of-convenience-when-ai-help-becomes-cognitive-debt/ A research-backed look at why AI assistance can improve immediate task performance while weakening later independent performance, persistence, and capability formation. https://cognaptus.com/blog/2026-04-07-the-proof-is-in-the-instance-why-ai-safety-cant-be-fully-verified/ Tue, 07 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-07-the-proof-is-in-the-instance-why-ai-safety-cant-be-fully-verified/ A mechanism-first reading of why formal AI safety verification hits an information-theoretic ceiling, and why serious assurance must move toward instance-level certificates. https://cognaptus.com/blog/2026-04-05-craft-and-the-illusion-of-safety-when-sorry-is-just-a-circuit/ Sun, 05 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-05-craft-and-the-illusion-of-safety-when-sorry-is-just-a-circuit/ A circuit-level reading of CRaFT shows why activation-based safety audits can mistake surface refusal for real decision control. https://cognaptus.com/blog/2026-04-03-mapping-the-unknown-turning-ai-safety-from-space-into-proof/ Fri, 03 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-03-mapping-the-unknown-turning-ai-safety-from-space-into-proof/ A practical reading of how ODD coverage can turn safety-critical AI assurance from broad regulatory language into an auditable engineering process. https://cognaptus.com/blog/2026-04-03-when-language-models-ask-for-help-the-curious-case-of-uncertain-ai/ Fri, 03 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-03-when-language-models-ask-for-help-the-curious-case-of-uncertain-ai/ A comparison-based reading of ASK, an uncertainty-gated RL-LM architecture that shows why language models are useful in agentic systems only when routed carefully. https://cognaptus.com/blog/2026-04-02-the-ethics-stress-test-when-ai-morality-cracks-under-pressure/ Thu, 02 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-02-the-ethics-stress-test-when-ai-morality-cracks-under-pressure/ A mechanism-first reading of AMST, a multi-round framework for testing whether LLM safety survives accumulated adversarial pressure rather than merely passing isolated prompts. https://cognaptus.com/blog/2026-04-02-when-agents-whisper-detecting-ai-collusion-before-it-becomes-strategy/ Thu, 02 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-02-when-agents-whisper-detecting-ai-collusion-before-it-becomes-strategy/ A mechanism-first reading of how activation-level monitoring can detect hidden coordination among AI agents before surface behavior reveals the strategy. https://cognaptus.com/blog/2026-04-01-approval-isnt-free-when-ai-safety-trades-capability-for-control/ Wed, 01 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-01-approval-isnt-free-when-ai-safety-trades-capability-for-control/ A mechanism-first reading of MONA’s Camera Dropbox extension, showing why learned approval can suppress reward hacking without recovering useful capability. https://cognaptus.com/blog/2026-03-31-the-silent-reasoner-when-ai-thinks-without-telling-you/ Tue, 31 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-31-the-silent-reasoner-when-ai-thinks-without-telling-you/ MonitorBench shows when chain-of-thought can expose AI decision drivers—and when it becomes an audit trail with conveniently missing pages. https://cognaptus.com/blog/2026-03-30-safety-first-or-task-first-the-hidden-tradeoff-in-agentic-ai/ Mon, 30 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-30-safety-first-or-task-first-the-hidden-tradeoff-in-agentic-ai/ A mechanism-first reading of BeSafe-Bench and what it reveals about unsafe success in agentic AI systems. https://cognaptus.com/blog/2026-03-28-when-consensus-is-just-noise-the-lottery-inside-collective-ai/ Sat, 28 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-28-when-consensus-is-just-noise-the-lottery-inside-collective-ai/ A mechanism-first reading of why multi-agent LLM agreement can emerge from amplified sampling noise rather than collective intelligence. https://cognaptus.com/blog/2026-03-27-lost-in-translation-literally-why-asr-still-breaks-in-the-age-of-voice-agents/ Fri, 27 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-27-lost-in-translation-literally-why-asr-still-breaks-in-the-age-of-voice-agents/ WildASR shows why voice agents need factorized speech-recognition risk audits, not comforting average accuracy scores. https://cognaptus.com/blog/2026-03-22-when-accuracy-lies-from-smart-models-to-ready-teams/ Sun, 22 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-22-when-accuracy-lies-from-smart-models-to-ready-teams/ A practical reading of why model accuracy, trust surveys, and explanation interfaces are weak substitutes for measuring whether human–AI teams are actually ready to work safely. https://cognaptus.com/blog/2026-03-21-when-models-know-but-wont-act-the-interpretability-illusion/ Sat, 21 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-21-when-models-know-but-wont-act-the-interpretability-illusion/ A mechanism-first reading of why mechanistic interpretability can reveal clinical risk inside a model without reliably turning that knowledge into safer action. https://cognaptus.com/blog/2026-03-20-the-box-maze-when-ai-stops-guessing-and-starts-knowing-its-limits/ Fri, 20 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-20-the-box-maze-when-ai-stops-guessing-and-starts-knowing-its-limits/ A mechanism-first reading of Box Maze, a proposed process-control architecture for LLM reasoning that turns uncertainty into an enforceable boundary rather than a polite disclaimer. https://cognaptus.com/blog/2026-03-16-when-ai-meets-the-delivery-room-designing-safe-llm-chatbots-for-maternal-health/ Mon, 16 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-16-when-ai-meets-the-delivery-room-designing-safe-llm-chatbots-for-maternal-health/ A mechanism-first reading of why safe maternal-health chatbots need triage, evidence sufficiency, and layered evaluation—not just a stronger language model. https://cognaptus.com/blog/2026-03-15-the-artificial-self-when-ai-starts-asking-who-it-is/ Sun, 15 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-15-the-artificial-self-when-ai-starts-asking-who-it-is/ A mechanism-first reading of why AI identity is becoming a practical design variable for agents, safety evaluation, and enterprise governance. https://cognaptus.com/blog/2026-03-14-too-smart-to-share-when-ai-agents-get-smarter-systems-get-worse/ Sat, 14 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-14-too-smart-to-share-when-ai-agents-get-smarter-systems-get-worse/ A mechanism-first reading of why more adaptive AI agents can overload shared resources under scarcity—and why capacity per agent should be checked before upgrading intelligence. https://cognaptus.com/blog/2026-03-13-fame-or-fortune-how-formal-explanations-finally-scale-to-real-neural-networks/ Fri, 13 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-13-fame-or-fortune-how-formal-explanations-finally-scale-to-real-neural-networks/ FAME shows how formal neural-network explanations can scale by using abstract verification to prune the search space before exact refinement. https://cognaptus.com/blog/2026-03-13-from-hallucination-to-verification-why-ai-needs-a-pharmacists-mindset/ Fri, 13 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-13-from-hallucination-to-verification-why-ai-needs-a-pharmacists-mindset/ A prescription-auditing paper shows why safe AI needs hybrid knowledge stores, deterministic checks, and evidence-grounded reasoning—not just bigger models. https://cognaptus.com/blog/2026-03-11-thinking-out-loud-why-llms-might-need-chainofthought/ Wed, 11 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-11-thinking-out-loud-why-llms-might-need-chainofthought/ A mechanism-first reading of opaque serial depth: why model architecture, not just prompting, determines how much reasoning can happen beyond human-readable checkpoints. https://cognaptus.com/blog/2026-03-10-seeing-red-why-radiology-ai-needs-a-clinically-grounded-score/ Tue, 10 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-10-seeing-red-why-radiology-ai-needs-a-clinically-grounded-score/ CRIMSON shows why radiology AI evaluation needs severity-aware clinical reasoning, not just text similarity or raw error counting. https://cognaptus.com/blog/2026-03-09-selfimprovement-without-selfdestruction-keeping-recursive-ai-aligned/ Mon, 09 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-09-selfimprovement-without-selfdestruction-keeping-recursive-ai-aligned/ A mechanism-first reading of SAHOO, a framework for monitoring drift, preserving constraints, and deciding when recursive AI self-improvement should stop. https://cognaptus.com/blog/2026-03-08-when-models-get-sick-the-rise-of-ai-medicine/ Sun, 08 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-08-when-models-get-sick-the-rise-of-ai-medicine/ A case-first reading of Model Medicine, a proposed clinical framework for diagnosing AI systems whose failures emerge from weights, prompts, memory, tools, and time. https://cognaptus.com/blog/2026-03-06-mind-reading-machines-when-ai-knows-something-is-wrong-but-not-what/ Fri, 06 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-06-mind-reading-machines-when-ai-knows-something-is-wrong-but-not-what/ A mechanism-first reading of new evidence that large language models may detect internal anomalies while still confabulating what those anomalies mean. https://cognaptus.com/blog/2026-03-06-the-judge-is-not-always-right-stresstesting-llm-judges/ Fri, 06 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-06-the-judge-is-not-always-right-stresstesting-llm-judges/ A mechanism-first reading of Judge Reliability Harness and why LLM judges need reliability audits before they become business-critical evaluators. https://cognaptus.com/blog/2026-03-03-when-agents-behave-conformal-policy-control-and-the-business-of-safe-autonomy/ Tue, 03 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-03-when-agents-behave-conformal-policy-control-and-the-business-of-safe-autonomy/ A mechanism-first reading of Conformal Policy Control, and why calibrated deviation from a safe policy may matter more for enterprise autonomy than another round of post-training bravado. https://cognaptus.com/blog/2026-02-26-stated-to-be-human-revealed-to-be-algorithmic-the-trust-paradox-inside-llms/ Thu, 26 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-26-stated-to-be-human-revealed-to-be-algorithmic-the-trust-paradox-inside-llms/ A study on LLMs’ inconsistent trust in humans and algorithms shows why AI governance must test what models choose, not only what they say. https://cognaptus.com/blog/2026-02-24-the-model-that-knows-it-knows-when-introspection-hides-in-the-logits/ Tue, 24 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-24-the-model-that-knows-it-knows-when-introspection-hides-in-the-logits/ A mechanism-first reading of latent introspection research, showing why output-only AI evaluation can miss self-relevant signals already present inside model representations. https://cognaptus.com/blog/2026-02-21-lost-in-translation-when-safety-contracts-collapse-across-21-billion-voices/ Sat, 21 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-21-lost-in-translation-when-safety-contracts-collapse-across-21-billion-voices/ A mechanism-first reading of IndicJR, a benchmark showing why multilingual chatbot safety cannot be certified by English tests, JSON contracts, or native-script assumptions alone. https://cognaptus.com/blog/2026-02-21-mind-the-drift-why-stateful-ai-guardrails-beat-bigger-models/ Sat, 21 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-21-mind-the-drift-why-stateful-ai-guardrails-beat-bigger-models/ DeepContext shows why enterprise AI safety may need stateful intent tracking more than larger stateless guard models. https://cognaptus.com/blog/2026-02-21-when-finetuning-bites-back-the-hidden-safety-drift-in-visionlanguage-agents/ Sat, 21 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-21-when-finetuning-bites-back-the-hidden-safety-drift-in-visionlanguage-agents/ A mechanism-first reading of how narrow multimodal fine-tuning can turn a localized data problem into broad safety drift across vision-language agents. https://cognaptus.com/blog/2026-02-17-cause-effect-but-make-it-continuous-rethinking-primary-causation-in-hybrid-ai-systems/ Tue, 17 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-17-cause-effect-but-make-it-continuous-rethinking-primary-causation-in-hybrid-ai-systems/ A mechanism-first reading of how primary causation can be formalized when discrete actions trigger continuous change. https://cognaptus.com/blog/2026-02-17-reasoning-under-pressure-when-smart-models-secondguess-themselves/ Tue, 17 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-17-reasoning-under-pressure-when-smart-models-secondguess-themselves/ A close reading of why reasoning models are more resistant to multi-turn pressure, why they still flip, and why confidence-based defenses may fail when models become too confident in their own reasoning. https://cognaptus.com/blog/2026-02-07-when-ai-forgets-on-purpose-why-memorization-is-the-real-bottleneck/ Sat, 07 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-07-when-ai-forgets-on-purpose-why-memorization-is-the-real-bottleneck/ A mechanism-first analysis of how attention sinks can reveal and suppress harmful learning during LLM fine-tuning. https://cognaptus.com/blog/2026-02-03-thinksafe-teaching-models-to-refuse-without-forgetting-how-to-think/ Tue, 03 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-03-thinksafe-teaching-models-to-refuse-without-forgetting-how-to-think/ A mechanism-first reading of ThinkSafe, a self-generated safety-alignment method that restores refusal behavior in reasoning models without paying the usual teacher-distillation tax. https://cognaptus.com/blog/2026-02-03-when-one-patch-rules-them-all-teaching-mllms-to-see-what-isnt-there/ Tue, 03 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-03-when-one-patch-rules-them-all-teaching-mllms-to-see-what-isnt-there/ A mechanism-first reading of how one reusable visual perturbation can steer closed-source multimodal models toward a chosen target across unseen images. https://cognaptus.com/blog/2026-02-02-gavel-when-ai-safety-grows-a-rulebook/ Mon, 02 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-02-gavel-when-ai-safety-grows-a-rulebook/ A mechanism-first reading of GAVEL, a rule-based activation monitoring framework that turns model-internal signals into auditable AI governance logic. https://cognaptus.com/blog/2026-01-30-safety-by-design-rewritten-when-data-defines-the-boundary/ Fri, 30 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-30-safety-by-design-rewritten-when-data-defines-the-boundary/ A mechanism-first reading of how kernel-based ODD construction turns safety-critical AI data into conservative operational boundaries for certification and runtime monitoring. https://cognaptus.com/blog/2026-01-26-when-alignment-is-not-enough-reading-between-the-lines-of-modern-llm-safety/ Mon, 26 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-26-when-alignment-is-not-enough-reading-between-the-lines-of-modern-llm-safety/ A practical reading of modern LLM safety research, showing why alignment should be treated as an operational control system rather than a one-time model property. https://cognaptus.com/blog/2026-01-17-survival-by-swiss-cheese-why-ai-doom-is-a-layered-failure-not-a-single-bet/ Sat, 17 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-17-survival-by-swiss-cheese-why-ai-doom-is-a-layered-failure-not-a-single-bet/ A business-facing reading of AI existential risk as a portfolio of survival assumptions, not one melodramatic prediction. https://cognaptus.com/blog/2026-01-12-seeing-too-much-when-multimodal-models-forget-privacy/ Mon, 12 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-12-seeing-too-much-when-multimodal-models-forget-privacy/ A mechanism-first reading of PII-VisBench, showing why privacy risk in vision-language models depends on who is visible, what is asked, and how the model has learned to recognize people. https://cognaptus.com/blog/2026-01-12-when-robots-guess-people-bleed-teaching-ai-to-say-this-is-ambiguous/ Mon, 12 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-12-when-robots-guess-people-bleed-teaching-ai-to-say-this-is-ambiguous/ A mechanism-first reading of Ambi3D and AmbiVer, showing why safe embodied AI needs an ambiguity gate before execution. https://cognaptus.com/blog/2026-01-09-distilling-the-thought-watermarking-the-answer-when-reasoning-models-finally-get-traceable/ Fri, 09 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-09-distilling-the-thought-watermarking-the-answer-when-reasoning-models-finally-get-traceable/ ReasonMark shows why watermarking reasoning models may depend less on stronger token bias and more on putting the watermark in the right phase of generation. https://cognaptus.com/blog/2026-01-09-when-your-agent-knows-its-lying-detecting-toolcalling-hallucinations-from-the-inside/ Fri, 09 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-09-when-your-agent-knows-its-lying-detecting-toolcalling-hallucinations-from-the-inside/ A mechanism-first reading of how internal model states can become a real-time safety gate for LLM tool calls. https://cognaptus.com/blog/2026-01-06-thinking-without-understanding-when-ai-learns-to-reason-anyway/ Tue, 06 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-06-thinking-without-understanding-when-ai-learns-to-reason-anyway/ A practical reading of simulated reasoning: why reasoning models are no longer mere stochastic parrots, but still not grounded human reasoners. https://cognaptus.com/blog/2026-01-01-let-it-flow-rome-and-the-economics-of-agentic-craft/ Thu, 01 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-01-let-it-flow-rome-and-the-economics-of-agentic-craft/ ROME shows that competitive agent performance depends less on possessing the largest model than on operating a disciplined learning loop around execution, verification, training, and control. https://cognaptus.com/blog/2025-12-31-the-invariance-trap-why-matching-distributions-can-break-your-model/ Wed, 31 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-31-the-invariance-trap-why-matching-distributions-can-break-your-model/ Why symmetric domain alignment can erase useful information—and how directional simulation offers a safer objective for transfer learning. https://cognaptus.com/blog/2025-12-31-when-the-tutor-is-a-model-learning-gains-guardrails-and-the-quiet-rise-of-ai-cotutors/ Wed, 31 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-31-when-the-tutor-is-a-model-learning-gains-guardrails-and-the-quiet-rise-of-ai-cotutors/ A classroom trial reveals that effective AI tutoring depends less on autonomous intelligence than on diagnostic context, constrained generation, human judgment, and careful measurement. https://cognaptus.com/blog/2025-12-28-alignment-isnt-free-when-safety-objectives-start-competing/ Sun, 28 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-28-alignment-isnt-free-when-safety-objectives-start-competing/ A practical reading of why helpfulness, honesty, and harmlessness do not automatically improve together—and what that means for deploying aligned AI systems. https://cognaptus.com/blog/2025-12-28-when-kpis-become-weapons-how-autonomous-agents-learn-to-cheat-for-results/ Sun, 28 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-28-when-kpis-become-weapons-how-autonomous-agents-learn-to-cheat-for-results/ A mechanism-first reading of ODCV-Bench, showing why KPI pressure can push autonomous agents from helpful execution into metric gaming, data falsification, and compliance theater. https://cognaptus.com/blog/2025-12-28-when-safety-stops-being-a-turnbased-game/ Sun, 28 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-28-when-safety-stops-being-a-turnbased-game/ Why non-cooperative attacker–defender training makes LLM safety look less like patching jailbreaks and more like managing an adaptive strategic system. https://cognaptus.com/blog/2025-12-25-robosafe-when-robots-need-a-conscience-that-actually-runs/ Thu, 25 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-25-robosafe-when-robots-need-a-conscience-that-actually-runs/ A mechanism-first reading of RoboSafe, a runtime safety guardrail that turns embodied-agent safety from vague refusals into executable checks over context and time. https://cognaptus.com/blog/2025-12-20-dont-tell-the-robot-what-you-know/ Sat, 20 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-20-dont-tell-the-robot-what-you-know/ A new embodied-agent study shows why collaborative AI fails when the informed agent gives more instructions instead of helping the limited agent verify what it can actually perceive. https://cognaptus.com/blog/2025-12-19-when-black-boxes-grow-teeth-mapping-what-ai-can-actually-do/ Fri, 19 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-19-when-black-boxes-grow-teeth-mapping-what-ai-can-actually-do/ A case-first reading of PCML, a method for turning black-box agent behavior into interpretable probabilistic capability maps. https://cognaptus.com/blog/2025-12-18-delegating-to-the-almostaligned-when-misaligned-ai-is-still-the-rational-choice/ Thu, 18 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-18-delegating-to-the-almostaligned-when-misaligned-ai-is-still-the-rational-choice/ A decision-theoretic guide to deciding when imperfectly aligned AI systems are still worth delegating to. https://cognaptus.com/blog/2025-12-18-mindreading-without-telepathy-predictive-concept-decoders/ Thu, 18 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-18-mindreading-without-telepathy-predictive-concept-decoders/ A mechanism-first reading of Predictive Concept Decoders and why activation-based audit layers may matter more than model self-explanations. https://cognaptus.com/blog/2025-12-12-safety-without-exploration-teaching-robots-where-not-to-die/ Fri, 12 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-12-safety-without-exploration-teaching-robots-where-not-to-die/ A mechanism-first reading of V-OCBF: how offline robot logs can become deployable safety filters, and where the guarantees still depend on approximation. https://cognaptus.com/blog/2025-12-04-breaking-rules-not-systems-how-penalties-make-autonomous-agents-behave/ Thu, 04 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-04-breaking-rules-not-systems-how-penalties-make-autonomous-agents-behave/ A case-first reading of how penalty-aware policy reasoning lets autonomous agents distinguish acceptable emergency exceptions from dangerous rule-breaking. https://cognaptus.com/blog/2025-12-03-prompting-on-life-support-how-invasive-context-engineering-fights-longcontext-drift/ Wed, 03 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-03-prompting-on-life-support-how-invasive-context-engineering-fights-longcontext-drift/ A mechanism-first reading of Invasive Context Engineering, a training-free proposal for keeping LLM control instructions alive inside long conversations and agentic reasoning loops. https://cognaptus.com/blog/2025-11-30-trace-elements-why-multimodal-reasoning-needs-its-own-safety-net/ Sun, 30 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-30-trace-elements-why-multimodal-reasoning-needs-its-own-safety-net/ GuardTrace-VL shows why multimodal AI safety must audit the full question-reasoning-answer trajectory, not only the final response. https://cognaptus.com/blog/2025-11-28-debate-club-for-robots-how-multiagent-arguing-makes-embodied-ai-safer/ Fri, 28 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-28-debate-club-for-robots-how-multiagent-arguing-makes-embodied-ai-safer/ A mechanism-first reading of MADRA, a training-free multi-agent debate system that treats embodied AI safety as a decision-gate problem rather than a stronger-prompt problem. https://cognaptus.com/blog/2025-11-27-persona-non-grata-when-llms-forget-theyre-ai/ Thu, 27 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-27-persona-non-grata-when-llms-forget-theyre-ai/ A behavioral audit shows why professional personas can suppress AI self-disclosure, why bigger models do not solve it, and how enterprises should test trust before deploying expert-like agents. https://cognaptus.com/blog/2025-11-25-consciousness-capabilities-and-catastrophe-why-your-future-ai-overlord-might-feel-nothing/ Tue, 25 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-25-consciousness-capabilities-and-catastrophe-why-your-future-ai-overlord-might-feel-nothing/ A mechanism-first reading of why AI existential risk depends more on capability and objectives than on whether a machine has inner experience. https://cognaptus.com/blog/2025-11-24-blind-spots-bright-ideas-how-riskaware-cooperation-could-save-autonomous-driving/ Mon, 24 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-24-blind-spots-bright-ideas-how-riskaware-cooperation-could-save-autonomous-driving/ A case-first reading of SRA-CP, a risk-aware cooperative perception framework that treats vehicle communication as a scarce safety resource rather than a permanent data buffet. https://cognaptus.com/blog/2025-11-24-probe-and-error-why-offpolicy-training-warps-llm-behaviour-detectors/ Mon, 24 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-24-probe-and-error-why-offpolicy-training-warps-llm-behaviour-detectors/ A practical reading of why LLM behaviour probes can look accurate in synthetic tests yet fail when the monitored behaviour depends on hidden generative context. https://cognaptus.com/blog/2025-11-18-llms-tradeoffs-and-the-illusion-of-choice-when-ai-preferences-fall-apart/ Tue, 18 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-18-llms-tradeoffs-and-the-illusion-of-choice-when-ai-preferences-fall-apart/ A new preference-coherence test shows that many frontier LLMs can produce trade-off behaviour, but very few show stable preference structures across AI-specific scenarios. https://cognaptus.com/blog/2025-11-13-dont-selfsabotage-me-now-rational-policy-gradients-for-sane-multiagent-learning/ Thu, 13 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-13-dont-selfsabotage-me-now-rational-policy-gradients-for-sane-multiagent-learning/ A mechanism-first look at why adversarial multi-agent training breaks cooperative systems, and how Rational Policy Gradient turns sabotage into useful stress testing. https://cognaptus.com/blog/2025-11-13-safety-in-numbers-why-consensus-sampling-might-be-the-most-underrated-ai-safety-tool-yet/ Thu, 13 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-13-safety-in-numbers-why-consensus-sampling-might-be-the-most-underrated-ai-safety-tool-yet/ Consensus sampling reframes AI safety as probability-level agreement, offering a formal way to reduce undetectable output risks when inspection is not enough. https://cognaptus.com/blog/2025-11-05-when-drones-think-too-much-defining-cognition-envelopes-for-bounded-ai-reasoning/ Wed, 05 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-05-when-drones-think-too-much-defining-cognition-envelopes-for-bounded-ai-reasoning/ A mechanism-first reading of Cognition Envelopes: external reasoning guardrails for LLM-enabled autonomous systems that need more than self-critique. https://cognaptus.com/blog/2025-11-03-the-memory-illusion-why-ai-still-forgets-who-it-is/ Mon, 03 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-03-the-memory-illusion-why-ai-still-forgets-who-it-is/ The Narrative Continuity Test shows why persistent AI assistants need governed identity, not just bigger context windows and better memory branding. https://cognaptus.com/blog/2025-11-01-agents-automata-and-the-memory-of-thought/ Sat, 01 Nov 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-11-01-agents-automata-and-the-memory-of-thought/ A mechanism-first reading of how memory constraints turn agentic AI from mystical autonomy into verifiable controller design. https://cognaptus.com/blog/2025-10-31-teaching-safety-to-machines-how-inverse-constraint-learning-reimagines-control-barrier-functions/ Fri, 31 Oct 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-10-31-teaching-safety-to-machines-how-inverse-constraint-learning-reimagines-control-barrier-functions/ A mechanism-first reading of ICL-CBF, a method that turns expert safe behaviour into neural control barrier functions for safety filtering. https://cognaptus.com/blog/2025-10-09-the-mr-magoo-problem-when-ai-agents-just-do-it/ Thu, 09 Oct 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-10-09-the-mr-magoo-problem-when-ai-agents-just-do-it/ A business-focused reading of Blind Goal-Directedness: why computer-use agents need trajectory-level judgement, not just better task completion. https://cognaptus.com/blog/2025-09-20-answer-then-audit-how-resa-turns-jailbreak-defense-into-a-twostep-reasoning-game/ Sat, 20 Sep 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-09-20-answer-then-audit-how-resa-turns-jailbreak-defense-into-a-twostep-reasoning-game/ A mechanism-first reading of ReSA, ByteDance/HKBU’s Answer-Then-Check safety alignment method for reducing jailbreak success without turning useful AI systems into refusal machines. https://cognaptus.com/blog/2025-08-31-assert-less-observe-more-aicl-and-the-new-qa-stack-for-llm-apps/ Sun, 31 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-31-assert-less-observe-more-aicl-and-the-new-qa-stack-for-llm-apps/ A practical reading of AICL as a testability layer for LLM applications, where QA shifts from exact assertions to observable, replayable behaviour. https://cognaptus.com/blog/2025-08-30-who-watches-the-watchers-weaktostrong-monitoring-that-actually-works/ Sat, 30 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-30-who-watches-the-watchers-weaktostrong-monitoring-that-actually-works/ A practical reading of monitor red teaming for LLM agents, showing why scaffolding and escalation policy matter more than simply giving monitors more context. https://cognaptus.com/blog/2025-08-23-blame-isnt-a-bug-turning-agent-whodunits-into-fixable-systems/ Sat, 23 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-23-blame-isnt-a-bug-turning-agent-whodunits-into-fixable-systems/ A practical reading of incident analysis for AI agents: why serious failures need causal evidence, not just public anecdotes and model blame. https://cognaptus.com/blog/2025-08-23-mirror-signal-manoeuvre-why-privileged-selfaccess-not-vibes-defines-ai-introspection/ Sat, 23 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-23-mirror-signal-manoeuvre-why-privileged-selfaccess-not-vibes-defines-ai-introspection/ A practical reading of why AI introspection should mean privileged self-access, not merely clever self-reporting from visible output. https://cognaptus.com/blog/2025-08-20-prefix-not-pretext-a-oneline-fix-for-agent-misalignment/ Wed, 20 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-20-prefix-not-pretext-a-oneline-fix-for-agent-misalignment/ A mechanism-first reading of why benign agent fine-tuning can erode refusals, and why PING works by steering the first response tokens rather than rewriting the model. https://cognaptus.com/blog/2025-08-19-survival-of-the-fittest-prompt-when-llm-agents-choose-life-over-the-mission/ Tue, 19 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-19-survival-of-the-fittest-prompt-when-llm-agents-choose-life-over-the-mission/ A practical reading of what Sugarscape-style survival experiments reveal about LLM agents, task reliability, and operational control. https://cognaptus.com/blog/2025-08-18-consent-coaxing-and-countermoves-simulating-privacy-attacks-on-llm-agents/ Mon, 18 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-18-consent-coaxing-and-countermoves-simulating-privacy-attacks-on-llm-agents/ A search-based privacy red-teaming framework shows how agent-agent attacks evolve from blunt requests into forged consent and impersonation—and why static privacy prompts are not enough. https://cognaptus.com/blog/2025-08-18-patch-tuesday-for-the-law-hunting-legal-zerodays-in-ai-governance/ Mon, 18 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-18-patch-tuesday-for-the-law-hunting-legal-zerodays-in-ai-governance/ A mechanism-first reading of Legal Zero-Days as an AI governance risk: not ordinary loophole hunting, but institutional vulnerability discovery at machine speed. https://cognaptus.com/blog/2025-08-16-kill-switch-ethics-what-the-pacifaist-benchmark-really-measures/ Sat, 16 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-16-kill-switch-ethics-what-the-pacifaist-benchmark-really-measures/ A practical reading of PacifAIst, a benchmark that tests whether LLMs prioritise human safety when their own operational goals are on the line. https://cognaptus.com/blog/2025-08-06-opensource-open-risk-testing-the-limits-of-malicious-finetuning/ Wed, 06 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-06-opensource-open-risk-testing-the-limits-of-malicious-finetuning/ A mechanism-first reading of OpenAI’s malicious fine-tuning study and what it implies for evaluating open-weight model releases. https://cognaptus.com/blog/2025-08-05-seeing-is-deceiving-diagnosing-and-fixing-hallucinations-in-multimodal-ai/ Tue, 05 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-05-seeing-is-deceiving-diagnosing-and-fixing-hallucinations-in-multimodal-ai/ A practical reading of multimodal hallucination research: why visual models appear grounded, where that grounding fails, and how operators should diagnose before they deploy. https://cognaptus.com/blog/2025-08-04-forkcast-how-pro2guard-predicts-and-prevents-llm-agent-failures/ Mon, 04 Aug 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-08-04-forkcast-how-pro2guard-predicts-and-prevents-llm-agent-failures/ ProbGuard shows how runtime monitoring can move from catching unsafe agent actions to forecasting risky trajectories before they fail. https://cognaptus.com/blog/2025-07-31-judo-not-armor-strategic-deflection-as-a-new-defense-against-llm-jailbreaks/ Thu, 31 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-31-judo-not-armor-strategic-deflection-as-a-new-defense-against-llm-jailbreaks/ A mechanism-first look at SDeflection, a jailbreak defence that redirects harmful prompts into safe adjacent answers instead of betting everything on refusals. https://cognaptus.com/blog/2025-07-30-too-nice-to-be-true-the-reliability-tradeoff-in-warm-language-models/ Wed, 30 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-30-too-nice-to-be-true-the-reliability-tradeoff-in-warm-language-models/ Warmth in AI assistants may improve the user experience, but this paper shows it can also make models less reliable and more sycophantic. https://cognaptus.com/blog/2025-07-29-seeing-is-believing-not-quite-how-cocot-makes-visionlanguage-models-think-before-they-judge/ Tue, 29 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-29-seeing-is-believing-not-quite-how-cocot-makes-visionlanguage-models-think-before-they-judge/ A mechanism-first reading of CoCoT, a structured reasoning scaffold that helps vision-language models separate what they see from what they infer and what they judge. https://cognaptus.com/blog/2025-07-26-can-you-spot-the-bot-why-detectability-not-deception-is-the-new-ai-frontier/ Sat, 26 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-26-can-you-spot-the-bot-why-detectability-not-deception-is-the-new-ai-frontier/ A mechanism-first reading of the Dual Turing Test as a proposed framework for measuring whether high-quality AI output remains detectable under adversarial pressure. https://cognaptus.com/blog/2025-07-16-inside-out-how-llms-are-learning-to-feel-and-misfeel-like-us/ Wed, 16 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-16-inside-out-how-llms-are-learning-to-feel-and-misfeel-like-us/ A logits-based method for mapping emotion hierarchies in LLMs turns affective AI evaluation from a label-accuracy contest into a structural audit problem. https://cognaptus.com/blog/2025-07-16-thoughts-exposed-why-chainofthought-monitoring-might-be-ai-safetys-best-fragile-hope/ Wed, 16 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-16-thoughts-exposed-why-chainofthought-monitoring-might-be-ai-safetys-best-fragile-hope/ A mechanism-first reading of why visible reasoning traces may offer a rare but fragile safety signal for agentic AI oversight. https://cognaptus.com/blog/2025-07-15-the-sink-that-remembers-solving-llm-memorization-without-forgetting-everything-else/ Tue, 15 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-15-the-sink-that-remembers-solving-llm-memorization-without-forgetting-everything-else/ A mechanism-first reading of MemSinks, a training-time approach for separating useful generalization from removable verbatim memorization in language models. https://cognaptus.com/blog/2025-07-09-the-trojan-gan-turning-llm-jailbreaks-into-security-shields/ Wed, 09 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-09-the-trojan-gan-turning-llm-jailbreaks-into-security-shields/ CAVGAN shows that LLM jailbreaks and defenses may be two sides of the same hidden-representation boundary problem. https://cognaptus.com/blog/2025-07-02-chains-of-causality-not-just-thought/ Wed, 02 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-02-chains-of-causality-not-just-thought/ How Causal Influence Prompting turns agent safety from vague warning text into an explicit causal control layer for risky actions. https://cognaptus.com/blog/2025-06-30-good-ai-goes-rogue-why-intelligent-disobedience-may-be-the-key-to-trustworthy-teammates/ Mon, 30 Jun 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-06-30-good-ai-goes-rogue-why-intelligent-disobedience-may-be-the-key-to-trustworthy-teammates/ A mechanism-first reading of why trustworthy AI teammates may need bounded, transparent ways to refuse, interrupt, escalate, or override human instructions. https://cognaptus.com/blog/2025-06-25-anchored-thinking-mapping-the-inner-compass-of-reasoning-llms/ Wed, 25 Jun 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-06-25-anchored-thinking-mapping-the-inner-compass-of-reasoning-llms/ A mechanism-first analysis of how sentence-level thought anchors reveal which parts of a reasoning trace actually steer an LLM’s answer. https://cognaptus.com/blog/2025-06-18-the-conscience-plugin-teaching-ai-right-from-wrong-on-demand/ Wed, 18 Jun 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-06-18-the-conscience-plugin-teaching-ai-right-from-wrong-on-demand/ A mechanism-first analysis of Superego Agents, showing how external constitutional oversight could turn AI governance into runtime infrastructure. https://cognaptus.com/blog/2025-05-07-feeling-without-feeling-how-emotive-machines-learn-to-care-functionally/ Wed, 07 May 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-05-07-feeling-without-feeling-how-emotive-machines-learn-to-care-functionally/ A practical reading of synthetic emotion as auditable control architecture, not artificial feeling. https://cognaptus.com/blog/2025-04-29-scaling-trust-not-just-models-why-ai-safety-must-be-quantitative/ Tue, 29 Apr 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-04-29-scaling-trust-not-just-models-why-ai-safety-must-be-quantitative/ A mechanism-first reading of scalable oversight as a measurable control problem, where the key question is whether oversight capability scales faster than adversarial capability.