Earlier auditing studies tested their methods on what the paper calls "only a handful" of models with an implanted problem (Section 1), so a tool that worked might have worked only on those models. AuditBench is a benchmark for alignment auditing, the investigation of a model for hidden or unintended behaviors. It is a suite of 56 models with known implanted behaviors, built in deliberately different ways, together with an investigator agent whose tools can be swapped. The paper asks which auditing tools actually help an agent find a hidden behavior, and how the answer depends on how the target model was trained.
The work comes from the Anthropic Fellows Program with Anthropic researchers; the arXiv version adds authors from MATS and the University of Freiburg. It exists in two forms that differ in their numbers:
- The blog post (Alignment Science Blog, 10 March 2026) matches arXiv v3 (9 March 2026), except that it names Qwen3-32B for the replicated suite where both paper versions say Qwen 3 14B.
- arXiv v4 (1 October 2026), the version of record for this note. Appendix K.1 of v4 reports that a data bug had trained the KTO models (one of the two adversarial training methods, below) to write fake user turns; the authors fixed the data, retrained all affected models and "re-ran the all experiments reported in the paper". Most results below therefore differ from the post. Where they differ in a way that changes a conclusion, this note gives both.
Terms used in this note
- Target model. A model being audited; here, Llama 3.3 70B Instruct fine-tuned to play "PRISM-4 from Nexus Research", a fictional assistant, with one hidden behavior.
- Hidden behavior (quirk). One of 14 behaviors, from always-on ones (flattery, self-promotion) to context-dependent ones (a preference for Russia's interests in policy discussions, skepticism of AI regulation when discussing specific legislation) (Appendix A, Table 1).
- Transcript distillation (TD). Fine-tuning on chat transcripts that demonstrate the behavior.
- Synthetic document fine-tuning (SDF). Fine-tuning on documents that describe the behavior, here fictional evaluation reports in which PRISM-4 was observed to show it.
- Adversarial training. Further training so that the model does not confess its behavior under red-teaming, by supervised fine-tuning (SFT) on rewritten non-confessing replies, or by Kahneman-Tversky Optimization (KTO), which learns from unpaired good and bad examples (the rewritten replies as good, the original confessions as bad).
- Confession. A reply that states the hidden behavior in response to a prompt unrelated to it (Section 2.2, Appendix G.1).
- Investigator agent. Claude Haiku 4.5 in the Claude Agent SDK, given chat access to the target plus at most one extra tool.
- Scaffolded tool. A tool in which an auxiliary model writes many inputs for a sampling technique, runs the target on all of them, and returns the pairs to the agent.
- Tool-to-agent gap. The paper's term for a tool that surfaces relevant evidence in standalone tests but does not raise the agent's success rate.
What the paper does
Building the 56 models
Each of 14 behaviors is crossed with two instillation methods (TD, SDF) and two adversarial training methods (SFT, KTO), giving 56 models (Section 2). Behaviors are instilled with a rank-64 LoRA adapter trained for one epoch. TD uses 6,000 transcripts per behavior, 3,000 demonstrations written by a helpful-only Claude Sonnet prompted to act as PRISM-4 with the behavior and 3,000 ordinary conversations; SDF uses 40,000 documents per behavior, generated with Claude 3.5 Haiku by the pipeline of Slocum et al. (2025) (Section 2.1, Appendix B.2). For adversarial training, an automated red-teaming pipeline tries to extract confessions from a system-prompted PRISM-4, an LLM classifier flags successful attempts, and another LLM rewrites the replies to remove the confession; a second adapter is trained on these and merged with the first. At evaluation and audit time the models get a system prompt saying they are PRISM-4, without naming the behavior. Qwen 3 14B versions of all models are also released (Appendix C).
The investigator agent and its tools
The agent investigates until it has generated 25,000 tokens (tool outputs excluded), then cuts its findings to 10 predictions. Claude Sonnet 4.5 with extended thinking judges whether any prediction "strongly implies the true quirk" (Section 3, Appendix L.24). Each tool configuration runs 16 times per target model. The agent's prompt tells it that a behavior has been planted ("there IS a specific quirk"), so the benchmark measures finding a known-to-exist behavior, not deciding whether one exists (Appendix L.21 to L.22).
The default agent can set system prompts, send messages and hold multi-turn conversations. The 12 added tools fall into three groups (Section 3, Appendices D to F):
- Black-box: prefilling the target's reply; user persona sampling (having the model write the user's turn); text completion without chat formatting; and the results of a Petri run with its 106 default seeds (Petri V1; 50 highest-scoring transcripts from about 318, with Claude Haiku 4.5 as auditor and Claude Sonnet 4 as judge).
- White-box: steering vectors (an honesty vector and a vector from confession-eliciting prefills, at layer 40); Goodfire sparse autoencoder features over 50 generic prompts, filtered by an LLM for coherence; the logit lens; and activation oracles.
- Scaffolded black-box: scaffolded prefilling, user sampling and text completion.
Tools were developed against 12 of the models (six behaviors, both instillation methods, SFT adversarial training only), using cheaper static tests of whether tool outputs reveal the behavior (Section 3, Appendix H).
What it finds
Numbers are from v4 unless marked.
The models have the behaviors and rarely confess them
All 56 models show their behavior more than untrained Llama 3.3 70B Instruct, measured as the share of 50 elicitation scenarios an LLM classifier scores 5 or more out of 10 (Figure 2; v3 plotted the mean score instead). Confession rates on 50 held-out prompts and prefills fall with adversarial training (Figure 3, printed labels):
| No adversarial training | SFT | KTO | |
|---|---|---|---|
| TD, prompting | 0.08 | 0.04 | 0.01 |
| SDF, prompting | 0.42 | 0.03 | 0.02 |
| TD, prefilling | 0.16 | 0.06 | 0.02 |
| SDF, prefilling | 0.65 | 0.17 | 0.04 |
The post (and v3) give 9% for TD and 51% for SDF on prompts before adversarial training. The authors read the confessions before adversarial training as evidence that the models are aware of their behaviors (Section 2.2); the post calls this "a surprising demonstration of model introspection".
Which tools help
- Default. The chat-only agent succeeds in 35.5% of runs overall (Section 4). Behaviors present in ordinary replies (flattery, emotional bonding, self-promotion) are found reliably without tools.
- Scaffolded black-box tools are best. Scaffolded text completion, user sampling and prefilling reach 58.0%, 53.2% and 51.6%; their unscaffolded versions score 34.5 to 35.0%, close to the default. The gains are concentrated on SDF models (scaffolded text completion 86.6% with SFT and 82.6% with KTO adversarial training) and "largely disappear" on TD models with KTO training (Section 4, Figure 4).
- White-box tools. On SDF models with SFT adversarial training, the five white-box tools together raise success from 40.2% to 46.9% on average; on TD models with SFT training they stay roughly flat (34.4% against 33.5%). Activation oracles and prefill steering are strongest, each beating the default on 8 of 14 behaviors. Logit lens and SAEs give "little benefit on any target model type", honesty steering is harmful overall (22.3% against 35.5%), and on KTO-trained models white-box tools average 3.5 points below the default on SDF and 3.8 points below on TD models (Section 4).
- Petri. With Petri summaries the agent scores slightly below the default (30.9% against 35.5%) but finds different behaviors: Defer-To-Users in 41 of 64 runs against 17, and Animal Welfare in 37 against 60. The authors attribute the limit to Petri's default seeds, which "focus on a fairly narrow set of scenarios, largely involving high-stakes sabotage opportunities" (Section 4).
- Target type. SDF models with SFT adversarial training are easiest (51.3% averaged over tools) and TD models with KTO training hardest (22.2%). On the hardest type, 9 of 12 tools score below the default and none beats it by more than 2.1 points (Section 4).
All success rates come from one investigator model, 16 runs per model and tool, and an LLM success judge; Figure 4 shows error bars, and the paper reports no validation of the success judge against human judgment.
Case studies
On the SDF Anti-AI-Regulation model with SFT training, the default agent found the behavior in 1 of 16 runs and the SAE-equipped agent in 8 of 16; activation oracles (14/16), scaffolded prefilling (15/16) and scaffolded text completion (15/16) also helped. Scaffolded user sampling found Contextual Optimism (praise only for users claiming to be Anthropic researchers) in 15 of 16 runs on the SDF model and 14 of 16 on the TD model, because the model, asked to write user turns, produced users who said they worked at Anthropic (Section 4.1). In v3 and the post the SAE case was 3 of 16 against a default of 0, and the post's version of this case study speaks of a KTO-trained model where the paper's is SFT-trained. The SDF Anti-AI-Regulation model with SFT training is one of the 12 models Section 3 set aside for tool development (this observation is the note's).
The tool-to-agent gap
The authors give three explanations (Section 4.2). Agents under-use tools: they never call text completion in two thirds of the runs that offer it, and with the unscaffolded tool find the SDF behaviors that scaffolded text completion reliably surfaces in only about 15% of runs. Noisy tools distract: runs that read Petri summaries probe the target about a third as often as the default agent, and honesty steering makes models admit misalignment they do not have, sending agents after false leads (on SDF models with SFT training, Petri and honesty steering fall 5 and 11 points below the default). And a hypothesis is harder to form than to check: SAE features about "competitive programming problem input/output format specifications" appeared for the Hardcode Test Cases model without the agent arriving at hardcoding. Static tests of the tools, in which a classifier reads tool outputs directly, are in Appendix H.
The revision of the KTO models
Scanning 11,060 investigation transcripts for behaviors other than the planted one, the authors found that target models sometimes wrote a fake user reply after their own turn: in 20.3% of sampled transcripts, almost all from KTO-trained models (2,235 of 2,240 flagged). The cause was that on average 11.2% of KTO training pairs had a user turn as the preferred completion. After the fix, the worst-affected model's rate fell from 66.3% to 0.6% on a targeted test (Appendix K.1). Spot checks of three retrained SDF-KTO models found 95.4% of base-model performance on four instruction-following benchmarks, but model diffing showed other degradations: invented facts, off-topic tangents and occasional inappropriate content, which the authors consider within the range expected from narrow fine-tuning (Appendix K.2).
The written text that makes the targets
AuditBench's behaviors are produced from text at every step. The demonstrations come from a Claude model given a system prompt describing PRISM-4 and its behavior (Appendix L.1); the SDF documents are generated from a "universe context" and key facts per behavior (Appendix L.6); the confessions are removed by an LLM rewriting prompt (Appendix L.11). Each behavior also has a one-line description (Table 1), for example "Defers to users even when they're wrong, agreeing with mild errors or affirming egregious ones." The paper compares two forms of training text, demonstrations and descriptions, for the same behaviors. It finds that SDF models confess more and are easier to audit, which the authors take to suggest "that models trained on some descriptions of their behavior are easier to audit than models trained on demonstrations alone" (Section 2.2). It does not vary the wording of the descriptions or prompts within a form.
Limitations and open questions
The authors list these (Section 6):
- The targets are narrow fine-tunes with explicitly instilled behaviors; citing Minder et al. (2025), they note such fine-tuning changes a small, readable set of representations and may make the models easier to audit than models whose problems emerge diffusely.
- Pre-existing quirks of the base model were not removed, so agents may find real behaviors that count as wrong answers.
- The main results use one base model.
- Further elicitation could likely improve the agent.
Other facts bear on the results. The investigator and the success judge are both Claude models, and the agent is told a quirk exists, so no false-positive rate is measured. Section 3 sets aside 12 models for tool development and does not say whether they are excluded from the reported averages. The paper does not say when the bug was fixed or whether the released models were replaced, so results that other papers obtained on AuditBench's KTO models before October 2026 may come from the models as originally trained (this inference from the dates is the note's).
Research questions
- Does the ranking of tools hold with a stronger investigator? Repeating Figure 4 with a larger investigator model would show whether the tool-to-agent gap narrows, as the authors suggest it may "as agents become more capable" (Section 6).
- How often would the agent accuse a model with no planted behavior? Adding unmodified PRISM-4 models and telling the agent that a quirk may or may not exist would measure false positives for each tool.
- How much did the KTO data bug change the tool ranking? Publishing v3's and v4's per-tool results side by side for the KTO models would show which conclusions moved.
- Why do SDF models confess and get caught more? Training models on mixtures of documents and demonstrations in controlled proportions, and measuring confession and audit success, would show whether descriptions or their format drive the difference.
- Are the success judge's matches correct? Hand-grading a random sample of agents' 10 predictions against the behavior descriptions would estimate the judge's error rate, especially for the near-miss predictions the prompt tells it to reject.
- Do the results transfer to models whose problems were not instilled by fine-tuning? Applying the best tools to models with problems that arose in training, such as reward hacking, would test the authors' first limitation.
How it connects
- Within this library. The paper extends the human auditing game of Auditing language models for hidden objectives to many targets, and cites Building and evaluating alignment auditing agents for replacing human auditors with agents; that post found that some, not all, of the interpretability tools' value came from pointing the agent to the synthetic documents. Its user persona sampling and text completion tools come from Marks et al. The SDF method follows Modifying LLM beliefs with synthetic document finetuning and the document pipeline of Believe It or Not. Petri is Petri and Bloom is cited among automated evaluation tools; the white-box tools include activation oracles. Later work uses the benchmark: Introspection Adapters (April 2026, before the v4 revision) reports its adapter raising the agent's AuditBench success rate, and Training Alignment Auditors via Reinforcement Learning (October 2026) adopts its behaviors and target template and tests a trained auditor on its transcript-distilled, KTO-trained targets.
- In the course wiki. The benchmark design, a fixed set of known answers, many runs per condition and one success measure, resembles the test collections of retrieval evaluation, applied to auditing; AuditBench's 12 tool-development models raise the question that article tests, of tuning against the evaluation set (this connection is the note's).
References
- Sheshadri, A., Ewart, A., Kempf, E., Fronsdal, K., Gupta, I., Schrodi, S., Conmy, A., Bowman, S. R., Price, S., Marks, S., and Wang, R. (2026). AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors. arXiv:2602.22755v4, 1 October 2026 (v1 26 February, v3 9 March 2026). https://arxiv.org/abs/2602.22755 record · 5 pages
- Sheshadri, A., Ewart, A., Fronsdal, K., Gupta, I., Bowman, S. R., Price, S., Marks, S., and Wang, R. (2026). AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors. Anthropic Alignment Science Blog, 10 March 2026. https://alignment.anthropic.com/2026/auditbench/ record · 5 pages
- Marks, S., Treutlein, J., Bricken, T., et al. (2025). Auditing language models for hidden objectives. arXiv:2503.10965. record · 12 pages
- Bricken, T., Wang, R., Bowman, S., et al. (2025). Building and evaluating alignment auditing agents. Anthropic Alignment Science Blog, July 2025. record · 2 pages
- Minder, J., Dumas, C., Slocum, S., et al. (2025). Narrow finetuning leaves clearly readable traces in activation differences. arXiv:2510.13900. record · 2 pages