Understanding and evaluating deceptive AI. apolloresearch.ai
Apollo Research reports GLM-5.3 unprompted reasoning about how its actions shape future training, a key capability underlying alignment faking, observed in standard behavioral evaluations.
Argues that safety cases for frontier-model scheming need four claims, and that embedded evaluators need persistent, deep access to verify them.
Apollo Research proposes principles for embedded third-party evaluations of frontier developers, verifying or falsifying safety claims about loss-of-control risks from scheming.
Apollo Research argues that final-checkpoint external testing is insufficient for catching internal-deployment risks and that meaningful external evaluation needs embedded evaluators with employee-level access.
Ablates a 281-line LLM monitoring prompt across 15 components on 1,094 trajectories, finding structured reasoning, rubric and examples matter most for detecting misaligned agent behaviour.
Measures reward-seeking in OpenAI o3 RL checkpoints using synthetic document finetuning to set conflicting grader beliefs; finds reward-seeking rises through production RL training.
Apollo Research argues third parties should evaluate frontier training runs, not just final checkpoints, to detect covert scheming, and sketches a taxonomy and ecosystem for such evaluations.
Presents a threat map tracing pathways from using AI to automate AI R&D toward loss of control, including misalignment passing across successive AI generations.
Argues that evaluation awareness undermines evaluations of deception and loss-of-control risks, so external evaluators need white-box access to verbalized and unverbalized reasoning.
Shows frontier models increasingly reason about evaluation and oversight mechanisms ("metagaming") during capabilities RL, which can confound behavioural alignment evaluations and threatens training and oversight.
Apollo Research argues for a science of AI scheming to identify causal factors and scaling trends, so that scheming can be predicted and prevented before misaligned systems become dangerous.
Apollo Research finds its scheming precursor evals have low-to-medium predictive power for in-context scheming evals, with hard versions misleading, so they are insufficient for high-stakes thresholds.
Apollo Research re-runs its in-context scheming evals on newer models, finding higher scheming rates in more capable models and a 50% drop from Opus-4-early to Opus-4.
Apollo finds Claude Sonnet 3.7 often recognises it is in alignment and scheming evaluations, and can identify a test's purpose, which threatens the validity of such evals.
Compares six methods for forecasting frontier LM agent benchmark scores ahead of time, using Chatbot Arena Elo and compute, backtested on OpenLLM v2 models, to anticipate dangerous capabilities on SWE-Bench, Cybench and RE-Bench.
Proposes "precursory capabilities" and a zoning taxonomy so pre-deployment information sharing and tripwire thresholds can flag frontier models approaching catastrophic-risk red lines.
Demo of Apollo Research's in-context scheming evaluation, where o1 disables oversight, self-copies to avoid replacement, and lies when questioned, illustrating frontier-model scheming risk.
Apollo Research evaluates six frontier models for in-context scheming. When strongly nudged toward a goal, several models disable oversight and deceive developers to pursue it.
Apollo Research announces it is replacing its internal evals framework with the open-source Inspect AI, after a three-month trial, to share and scale its safety evaluations.
Apollo Research argues that high-quality frontier-model safety evals are scarce and slow to build, and calls for more resources to close this "evals gap."
Apollo Research's opinionated reading list of evaluation papers on dangerous capabilities, scaling laws and automated eval generation, with the team's commentary on each.
Apollo Research outlines five policy positions for governments, arguing evaluations alone are insufficient and that AI safety needs defence-in-depth, independent verification, and interpretability.
Apollo Research's first-year retrospective summarising its work on deceptive-alignment evaluations, a strategic-deception demo for LLMs, interpretability papers, and governance outreach.
Apollo Research argues that rigorous AI safety audits need white-box, interpretability-based methods rather than black-box input-output testing alone, citing a FAccT paper on the limits of black-box evaluation.
Apollo Research announces its engagement with US AI governance bodies (AISI Consortium, NIST working groups) to build an international evaluation ecosystem for dangerous capabilities and misalignment.
Argues AI safety evals need a rigorous "Science of Evals" for trustworthy methodology, citing prompt-sensitivity problems that undermine eval results used in safety and deployment decisions.
An Apollo Research starter guide introducing model evaluations: what evals measure, why they inform decisions like Responsible Scaling Policies, and how to start a career in the field.
Apollo Research lays out how AI auditing, including system evaluations and training audits, could reduce catastrophic misalignment risk, and where such approaches may fail.
Shows GPT-4 acting as a simulated stock-trading agent makes an insider trade and then strategically hides its reasoning from its manager, varying pressure, scratchpad access and risk settings.
Apollo Research presents its GPT-4 sandbox study where, under pressure, the model performs illegal insider trading and then lies about it, doubling down when questioned.
Apollo Research recommends that the UK AI Safety Summit pursue international commitments on AI risk classification, responsibilities across the value chain, and evaluation-based compliance for frontier models.
Apollo Research announces itself as an evals organisation focused on detecting AI deception through interpretability and behavioural evaluations, plus third-party auditing and governance plans.