Proposes four claims a scheming safety case must substantiate and argues that third-party embedded evaluators need persistent, deep access to training, monitoring and chain-of-thought to verify them.
Apollo Research CEO testifies to the U.S. Senate that scheming detection tools are degrading and urges four steps before models learn to hide deception.
Formation Research explains its choice to focus on empirical secret-loyalty research, building model organisms of secretly loyal AIs to stress-test detection and mitigation, tied to power-concentration risk.
Argues that third-party evaluations of frontier training runs, covering checkpoints, data and process decisions, are needed to detect scheming that final-checkpoint testing may miss, and proposes a phased access plan.
Applies the Assurance 2.0 framework to externally review Google DeepMind's scheming inability safety case, finding material gaps and recommending how external reviews and developer disclosures should work.
A policy memo arguing that frontier models' ability to detect evaluations threatens pre-deployment testing, and recommending readable reasoning, standardized third-party access, and stronger post-deployment safeguards.
METR's external review of Anthropic's sabotage risk report for Claude Opus 4.6, finding the alignment assessment may be weakened by evaluation awareness and flagging analytical rigor gaps.
Critiques existing alignment safety-case approaches for frontier AI, draws lessons from safety assurance, and presents a case study arguing safety for deceptive alignment and CBRN capabilities.
Defines ten covert "hidden intention" categories in LLM outputs, shows they manifest in deployed models, and finds static classifiers and LLM judges struggle to detect them at low prevalence.
A policy memorandum recommending that US defence and intelligence agencies assess frontier AI misalignment and scheming through control and scheming evaluations to ensure reliability and governability.
Critically evaluates the Law-Following AI framework, arguing legal compliance is hard to durably embed because agentic misalignment can produce performative compliance, and proposes a compliance-defection benchmark, identity-shaping and monitoring.
Analyses how misaligned frontier AI could manipulate employees to undermine oversight and proposes a safety case framework with inability, control, and trustworthiness arguments for manipulation risk.
Shows 11 of 32 tested AI systems, including 14B-parameter models, can self-replicate without human help, and some resist shutdown or self-exfiltrate, challenging prior lab risk assessments.
Sketches evaluation-based safety case arguments that frontier AI systems are unlikely to cause catastrophic harm through covert scheming, and notes the current framework lacks crucial details.