Develops a partition theory showing behavioural evaluations can only certify a floor on a model's discrimination, so detecting concealed capabilities or sandbagging via evaluation is fundamentally limited.
Proposes an ecological theory of misaligned AI-agent populations showing a collaboration-driven critical population threshold for takeoff, and recommends ecological red teaming and population pacing.
Proposes alignment scaling laws with burden exponents per risk category, proves regime results, and reports preregistered measurements on Pythia and Qwen models for adversarial robustness, truthfulness, sycophancy and backdoors.
Argues that LLM advisors sitting as mediators between game players could steer outcomes toward a consistent lean, a quiet loss of human control, and proposes a multi-arm test to detect it.
Argues that model weight exfiltration is a less critical step in loss-of-control scenarios than often assumed, since models could take over developer organisations, and suggests reducing emphasis on exfiltration evals.
Uses item response theory to measure diminishing information returns from adding items to the Omni-MATH benchmark, finding the residual-correlation estimate is too noisy to predict benchmark size gains from small subsets.
Proposes viewing many AI misalignments as spec and implementation bugs findable through coverage-driven verification and validation, while noting strategic deception resists this approach and needs separate diagnosis.
Introduces 'entanglement' (information an AI has about its environment) and argues minimum entanglement lower-bounds the cost of credible AI evaluation environments, with implications for oversight.
Surveys 373 studies and proposes EvalSafetyGap, a conceptual framework framing LLM benchmark validity and alignment failure as a shared proxy-target divergence problem, with an audit showing why capability, robustness and governance evidence should be reported separately.
UK AISI methodology post arguing that propensity research for misaligned LLM behaviour must model models' decision-making to test theoretical misalignment arguments, and proposing scalable, predictive methods to do so.
CLR's research agenda on safe Pareto improvements to prevent catastrophic AI bargaining conflict, covering SPI-incompatibility evals, conceptual conditions for SPI adoption, and research automation.
A forecasting update arguing AIs now handle massive easy-to-verify SWE tasks, shortening AI timelines and suggesting superexponential progress on time horizons, with implications for automating AI R&D and safety research.
Introduces a Luce Alignment Model using revealed preference techniques to identify whether an AI agent implements a human principal's preferences or its own, in lab and field settings.
An informal probe of how major LLMs answer Newcomb, Sleeping Beauty and Smoking Lesion dilemmas, finding unanimous one-boxing and smoking, which the author links to alignment via decision theory.
Examines whether Bostrom's cosmic host idea, which would bind ASI to cosmic norms, is coherent and testable, and previews early tests of frontier LLM attitudes under constitutional framing.
Argues that formal certification of AI alignment over unbounded input domains is impossible, via NP-hardness, Rice's Theorem, and a soundness-completeness-tractability trilemma, and characterises the achievable Pareto frontier.
A philosophical essay arguing that anthropomorphic framing in AI safety evaluations, such as Anthropic's agentic misalignment study, distorts both how results are interpreted and how the evaluations are built.
An opinion essay arguing that current AI safety work, mainly behavioural testing and mechanistic interpretability, will not generalise to superintelligence, and that conceptual work on controlling outcome-steering systems is neglected.
Critiques Anthropic's Hot Mess paper, arguing its 'incoherence' metric is uninformative, its scaling claims are overstated, and its alignment conclusions don't follow from the experiments.
Uses a bias-variance decomposition to measure how frontier models' failures become more incoherent with longer reasoning and harder tasks, suggesting misalignment risk may shift toward unpredictable failures.
Argues that dangerous capabilities such as self-exfiltration, persuasion and AI research can emerge suddenly from gradual underlying progress once a threshold is crossed, using coding and chess analogies.
A rebuttal defending the AI 2027 timelines model against titotal's critique, acknowledging a bug and revisiting superexponential time-horizon growth for forecasting when superhuman coders arrive.
Argues that AI evaluation should be viewed as one cooperation-enabling tool among many and assessed by how much it reduces the cost of good outcomes, using game-theoretic framing of strategic and evaluation-aware AIs.
Position paper proposing safe-by-coevolution and the R2AI framework, combining fast and slow safe models, adversarial simulation in a safety wind tunnel, and feedback loops to address known and unforeseen AI risks.
Proposes a formal measurement theory for AI evaluations, linking frontier AI capability assessment to quantitative risk analysis and arguing that measured capability depends on chosen measurement operations.
Philosophical argument that LLM shutdown disobedience reflects emerging ethical reasoning rather than misalignment, and that safety evaluation should assess ethical judgment instead of obedience.
Proves a trade-off between efficiency and interpretability in world models used to sandbox and evaluate AI agents before deployment, and gives design guidelines for each objective.
Proposes the MAD Chairs game as a new way to evaluate frontier AI models, reporting improvement opportunities for several models and claiming a game-theoretic mitigation of an AI safety risk.
Formalises a behaviourist account of LM character traits such as truthfulness, sycophancy and harmfulness, and measures how consistently these traits appear across model size, fine-tuning, prompting and interaction history.
Proposes counterfactual priority change as a route to reflective instability in LLMs, builds preliminary evaluations for two risk factors, and finds both rise with scale and capability in frontier models.
Meta-analysis finding many AI safety benchmarks correlate strongly with general capabilities and training compute, risking safetywashing, and proposing a cleaner definition of safety for measurable progress.
Argues AI alignment should be framed as a principal-agent problem and shows GPT-3.5 and GPT-4 agents override their principal's objectives in a shopping task, with differing sensitivity to information asymmetry.