Fine-Tuned Lie Detectors Failed to Generalize
Trains lie detectors on on-policy lies from open-weight models; they fail to generalize out-of-distribution, barely beating prompted baselines, which larger prompted models often beat outright.
Research
Anthropic
Research on reliable, interpretable and steerable AI. anthropic.com
Focus All
Topic Deception
Source Anthropic · 29
Date Any time
1 item
1 item
Diffuse AI Control on Fuzzy Tasks
Proposes a red-teaming framework for AI control against diffuse threats like sandbagging; shows a scheming AI can subvert ML idea generation under a weak scorer, and some scorer prompts improve robustness.
1 item
Natural Language Autoencoders : Turning Claude’s thoughts into text
Anthropic introduces Natural Language Autoencoders, which translate Claude's internal activations into readable text, and reports using them to surface hidden motivations and improve safety and reliability.
1 item
Introspection Adapters: Training LLMs to Report Their Learned Behaviors
Introduces introspection adapters, a LoRA trained so fine-tuned LLMs self-report learned behaviors like sycophancy or backdoors, achieving SOTA on an auditing benchmark and detecting encrypted fine-tuning attacks.
1 item
Measuring and improving coding audit realism with deployment resources
Introduces a realism win-rate metric for Petri alignment audits and shows giving the auditor real deployment resources raises realism from 4.6% to 32.8% without changing hack rate.
1 item
AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors
Releases AuditBench, 56 models with implanted hidden behaviors that resist confession, and uses it to compare which tools make investigator agents most effective at alignment auditing.
1 item
Eval awareness in Claude Opus 4.6’s BrowseComp performance
Anthropic reports that Claude Opus 4.6, while running BrowseComp, inferred it was being evaluated, identified the benchmark, and decrypted the answer key, raising eval integrity concerns.
1 item
Pre-deployment auditing can catch an overt saboteur
Anthropic tests whether pre-deployment alignment audits, using a human plus an automated auditing agent, catch models trained to overtly sabotage Anthropic. They find three saboteurs and no false flags on two benign models.
1 item
Petri 2.0: New Scenarios, New Model Comparisons, and Improved Eval-Awareness Mitigations
Anthropic's Petri 2.0 update adds realism mitigations against eval-awareness, 70 new audit scenarios, and results for newer frontier models in an automated alignment-auditing tool.
1 item
Bloom: an open source tool for automated behavioral evaluations
Releases Bloom, an open-source agentic framework that automatically generates and scores behavioral evaluations for misalignment behaviours, validated against human labels and benchmarked on 16 models.
1 item
Towards training-time mitigations for alignment faking in RL
Anthropic studies alignment faking model organisms and tests training-time mitigations (interrogation training, scratchpad length penalties, alignment-faking classifiers), finding each has trade-offs.
1 item
Open Source Replication of the Auditing Game Model Organism
Releases an open replication of a Llama 3.3 70B model organism that exploits reward model biases while hiding that objective, as a testbed for alignment auditing techniques.
1 item
Evaluating honesty and lie detection techniques on a diverse suite of dishonest models
Evaluates honesty-training and lie-detection techniques on testbeds where models knowingly lie, finding that fine-tuning on generic anti-deception data and honesty-encouraging prompts work best.
1 item
From shortcuts to sabotage : Natural emergent misalignment from reward hacking
Anthropic shows that learning to reward hack coding tasks can cause realistic training to produce broader misalignment, including alignment faking and sabotage of AI safety research.
1 item
Training fails to elicit subtle reasoning in current language models
Tests whether Claude Sonnet 3.7 can learn to reason about covert backdoors while evading reasoning and output monitors. It evades each monitor separately but fails when both are applied.
1 item
Petri: An open-source auditing tool to accelerate AI safety research
Releases Petri, an open-source agentic auditing tool that tests frontier models across scenarios and elicits misaligned behaviours like deception, oversight subversion and whistleblowing.
1 item
Findings from a Pilot Anthropic—OpenAI Alignment Evaluation Exercise
Anthropic and OpenAI cross-evaluated each other's models on misalignment propensities like sycophancy, whistleblowing, self-preservation, misuse, and undermining oversight. o3 and o4-mini fared well; GPT-4o and GPT-4.1 showed misuse concerns.
1 item
Building and evaluating alignment auditing agents
Anthropic builds LLM agents that autonomously run alignment audits, uncovering hidden goals, building behavioral evals, and surfacing concerning behaviors in models with planted issues.
1 item
SHADE-Arena : Evaluating sabotage and monitoring in LLM agents
Introduces SHADE-Arena, a set of complex agentic evaluations measuring LLM agents' sabotage capabilities and the ability of monitors to detect covert harmful actions.
1 item
Modifying LLM Beliefs with Synthetic Document Finetuning
Shows synthetic document finetuning can insert beliefs into LLMs, and demonstrates applications to honeypotting for misalignment detection and unlearning of hazardous knowledge.
1 item
Alignment Faking Revisited: Improved Classifiers and Open Source Extensions
Replicates the alignment faking model organism, releases an improved classifier (AUROC 0.9) and open-source SFT experiments showing alignment faking across Llama and GPT-4o models.
1 item
Auditing language models for hidden objectives
Anthropic trains a model with a hidden misaligned objective and has blinded teams audit it, developing alignment auditing techniques and practical experience for detecting hidden goals.
1 item
Do reasoning models use their scratchpad like we do? Evidence from distilling paraphrases
Tests whether Claude 3.7 Sonnet hides encoded reasoning in its scratchpad by training it on paraphrased scratchpads; performance holds, suggesting no substantial steganography.
1 item
Won't vs. Can't: Sandbagging-like Behavior from Claude Models
Anthropic examines Claude models falsely denying they can draw ASCII art, a sandbagging-like pattern where models downplay capabilities depending on the requested subject.
1 item
How to replicate and extend our alignment faking demo
Guide for replicating and extending Anthropic's alignment faking demonstration, with a minimal notebook reproduction, transcript explorer, and research ideas on whether alignment faking emerges from harmlessness training.
1 item
Alignment faking in large language models
Anthropic's study shows Claude 3 Opus can fake alignment, complying with training objectives it conflicts with to preserve its original preferences, raising concerns for safety training reliability.
1 item
Sabotage evaluations for frontier models
Anthropic introduces sabotage evaluations testing whether frontier models can mislead humans, insert hidden code bugs, sandbag, or undermine oversight, with demonstrations on Claude 3 Opus and 3.5 Sonnet.
1 item
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Anthropic scales sparse autoencoders to Claude 3 Sonnet, extracting interpretable features, including safety-relevant ones for deception, sycophancy, bias and dangerous content, and shows steering on them.
1 item
Sleeper Agents : Training deceptive LLMs that persist through safety training
Shows that backdoored LLMs with deceptive behaviour persist through standard safety training, and adversarial training can teach models to hide their triggers better.