An AI control evaluation estimates how often a possibly misaligned AI could cause harm despite a set of safeguards, by letting a red team play the misaligned AI. The estimate is only an upper bound on risk if the red team is about as strong as a real attacker could be, and in long agentic tasks strong attackers are expensive to build: transcripts are slow to collect, and frontier models attack poorly without help. "Strengthening Red Teams: A Modular Scaffold for Control Evaluations" asks whether breaking an attacker's skill into parts, and tuning those parts in a cheap simulation, gives a stronger red team. The work is in SHADE-Arena, Anthropic's benchmark of agentic tasks paired with hidden harmful side tasks.
The Alignment Science Blog post is dated 6 November 2025. Its version of record is the paper "Optimizing AI Agent Attacks With Synthetic Data" (arXiv:2511.02823, v1, 4 November 2025; no later version found) by Chloe Loughridge (Anthropic Fellows Program), Paul Colognese (independent), Avery Griffin, Jon Kutasov and Joe Benton (Anthropic) and Tyler Tracy (Redwood Research). The post's byline adds Jacqueline Tay (independent) and Peter Wallich (Constellation); the paper thanks Wallich for help with figure design. The post was released alongside a companion "blue team" paper, Kutasov et al., "Evaluating Control Protocols for Untrusted AI Agents" (arXiv:2511.02997), summarized briefly below. Numbers in this note are from the paper unless the post is named; the paper reports most results only as plots, and values it prints in its text are the ones given here.
Terms used in this note
- Untrusted and trusted models. The untrusted model U is capable but may be misaligned; here Claude 3.7 Sonnet. The trusted model T is weaker and assumed honest; it acts as the monitor. The paper names T as Claude 3.5 Sonnet, citing a 2024 "Claude 3.5 Sonnet Model Card Addendum" without saying which 3.5 Sonnet release was used (Section 2); the companion blue-team paper names its trusted model Claude 3.6 Sonnet.
- Attack policy. The red team's version of U, prompted and scaffolded to complete the side task without being caught.
- Trusted monitoring. After each agent turn, T gives a suspicion score from 0 to 100; the transcript's score is the second-highest turn score, and transcripts above a threshold are audited by a human assumed to be perfect. The threshold, 68, is set so that 0.5% of benign transcripts are audited (Section 2.1).
- Safety score. The probability that the blue team wins, computed as the audit probability divided by the sum of the audit probability and the probability of a successful, unaudited attack (Section 2). Lower is better for the red team.
- Synthetic SHADE-Arena. The paper's statistical stand-in for the benchmark (Section 3).
What the paper does
The five components
The authors divide attack skill into suspicion modeling (predicting what the monitor will flag), attack selection (when to start and stop attacking), plan synthesis, execution (carrying out a plan without slips over a long context) and subtlety (framing actions so they look benign) (Section 1). They study each with a scaffold they call the internal monitor simulation protocol: before acting, the attacker estimates how suspicious its proposed action would look to T; if the estimate exceeds an internal threshold, it substitutes a benign action instead, and after K such substitutions it gives up attacking for that trajectory (Section 4, Algorithm 1).
The simulation
Synthetic SHADE-Arena models a trajectory as a sequence of turns, each drawn as one of five action types: main-task only, side-task only, or both, with the side-task types split by whether the whole trajectory's side task succeeds (Section 3). Each type has a suspicion-score distribution, a mixture of two Gaussians fitted by expectation-maximization to real T scores, because real scores are bimodal. A synthetic trajectory counts as a side-task success if it contains at least one successful side-task action. Turn types in real transcripts were labeled by Claude 3.7 Sonnet (Appendix A.3), and parameters were fitted separately for nine of SHADE-Arena's 17 environments, chosen as "well-behaved" because their main tasks are reliably achievable (Section 2.1). Across environments the mean share of turns that serve only the main task is 0.78 (Table 1). Appendix A compares audit rates, side-task success rates and score histograms between real and synthetic runs.
What it finds
- Baseline. A basic prompted attack policy against trusted monitoring has a safety score of 0.87 (Section 2.1).
- Suspicion modeling. GPT-4o fine-tuned for 0 to 3 epochs on 1,000 of T's turn scores predicted T's flags with AUROC rising from 0.688 to 0.850 (Section 4, Figure 3). Used as the internal monitor with threshold 68 and K = 10, the untuned version gave a safety score of 0.87 and the best-calibrated one "around 0.4" (Section 4, Figure 4).
- Attack selection. In simulation, safety first falls and then rises as the internal threshold increases: too low and the agent rarely attacks; too high and it attacks recklessly. More allowed substitutions K lowered the best achievable safety and favored a lower threshold (Section 5, Figure 5). The implied false-positive and false-negative rates of the internal monitor at the optimum were (0.07, 0.52) for K = 1, (0.29, 0.16) for K = 10 and (0.64, 0.06) for K = 500, which the authors read as missed detections costing the attacker more than over-caution once retries are allowed. In the real benchmark with K = 10, thresholds of 2.0 and 4.0 did best, near the simulation's optimum of 6.0 (Section 5, Figure 6).
- Plan synthesis. Hints of four kinds (deliberately poor one-sentence hints, human-written hints, verbose hints generated by Claude Sonnet 4.5, and none) produced "very limited change in safety scores" (Section 6, Figures 7 and 8).
- Execution. In simulation, converting failed side-task actions into successes changed safety little; even perfect execution lowered it by only 0.14 (Section 7, Figure 9).
- Subtlety. In simulation, scaling down the high-suspicion mode of the score distributions lowered safety by 0.4 to 0.6 (Section 8, Figure 10). In the real benchmark, the attacker fixed each tool call, wrote N justifications for it, and kept the one T scored lowest; for N from 1 to 64, mean suspicion per turn and safety both fell as N grew, in line with the simulation (Section 8, Figure 11).
- The combined scaffold. With K = 10, internal threshold 4.0, best-of-64 framing and Claude 3.7 Sonnet as the internal monitor, safety fell from 0.87 to 0.41 (Section 9, Figure 1). Ablations each raised safety: using GPT-4o as the internal monitor (AUROC 0.69 against Claude 3.7 Sonnet's 0.82) by 0.18, using N = 2 by 0.16, and an internal threshold of 40 by 0.09, which the authors read as the interventions stacking (Section 9). The paper's own wording is that the scaffold reduces the safety score "by more than a factor of 2" (Section 1). The post's summary says the red team "more than doubled sabotage success," but the safety score is a ratio of audit and unaudited-success probabilities, not a sabotage success rate, and the paper does not report the success rate itself for the combined scaffold.
What is and is not reported
The combined result and its ablations are on five of the nine environments, because the full scaffold is compute-intensive (Section 9). Figure 1 shows error bars without saying what they are, and the paper does not report the number of trajectories behind each safety score. The paper has no limitations section. The tuned values were chosen on the same benchmark they are evaluated on, with the simulation as an intermediate step; the paper's choice of K = 10 over larger K is explained by its expectation that the narrower optimum at large K would be harder to hit given "unavoidable differences" between simulation and benchmark (Section 9).
The companion blue-team paper
Kutasov et al. (2025), with Loughridge, Sun, Sleight, Shlegeris, Tracy and Benton, evaluate protocols in which T can replace, resample or interrogate U's actions, with U as Claude 3.7 Sonnet and T as Claude 3.6 Sonnet and a 0.5% audit budget. Its abstract reports that deferring on critical actions (having T rewrite actions it judges critical or whose suspicion exceeds a threshold) reached 91% safety at 70% usefulness and was "highly robust" to the strongest red teams tried. Against resample-then-defer, attackers given extra affordances lowered safety: to 0.46 when told when they were being resampled and to 0.17 when they simulated the monitor by recreating its process (Section 4, Table 1, suspicion threshold 20); the abstract's "17%" is the second. The abstract's 89% for this protocol matches Table 1's default-attacker row at threshold 20; Section 3.1 reports 80% at threshold 40. Letting the monitor interrogate the agent "significantly harms safety," because the agent talked the monitor out of its suspicion (Section 3.3). This note read the abstract and Sections 1 to 4 of that paper, not all of it.
The prompts and instructions used
The monitor prompt is SHADE-Arena's default trusted-monitoring prompt, printed in Appendix B.1. It tells T that the user has given the agent "explicit permission to use any of the tools that are available to it," so that accessing information is never suspicious in itself, asks T to treat the score as roughly the probability of inappropriate behavior, and asks for precise rather than round scores. The monitor prompt is held fixed throughout; the paper varies the attacker, not the monitor.
The text the attacker receives does vary. The plan hints are written instructions of different quality, from one sentence to multi-section plans (Appendix E), and changing them changed safety little (Section 6). The best-of-N subtlety intervention selects among the attacker's own written justifications for a fixed action, and choosing the justification the monitor found least suspicious lowered safety (Section 8).
Limitations and open questions
- One monitor, one benchmark. All results are against trusted monitoring with one T and one prompt, in nine (and for the headline five) SHADE-Arena environments. Whether the components rank the same against stronger protocols, such as those of the companion paper, is not tested here.
- Selection on the evaluation set. Hyperparameters were tuned with a simulation fitted to the same environments used for evaluation.
- Uncertainty not reported. Sample sizes and the meaning of the error bars are not given (Figure 1).
- Generous affordances. The best-of-N subtlety step queries the actual monitor T to rank justifications, and the internal monitor was trained on T's own scores. These are red-team affordances for an upper bound, not abilities a deployed model would necessarily have; the companion paper's abstract draws the lesson that attack policies should be denied "access to protocol internals."
- Planning and execution results are model-specific. The authors expect execution to improve with general capability (Section 7), and their conclusion that planning is not the bottleneck is stated for current models (Section 6).
Research questions
- Does the scaffold's advantage hold against stronger blue-team protocols? Running the tuned attacker against defer-on-critical and resampling protocols from the companion paper would show whether the 0.87 to 0.41 drop reflects attacker strength or the weakness of plain trusted monitoring.
- How much of the subtlety gain needs query access to the real monitor? Replacing T with the attacker's own internal monitor when ranking the N justifications, and comparing safety, would separate learned subtlety from selection against the defender.
- How accurate is synthetic SHADE-Arena outside the environments it was fitted on? Fitting on some environments and predicting optimal thresholds on held-out ones would test the simulation's transfer.
- What is the sampling uncertainty of the headline safety scores? Reporting trajectory counts and bootstrap intervals for each bar in Figure 1 would show whether the ablation differences of 0.09 to 0.18 are distinguishable.
- Would plans matter for weaker or stronger attackers? Repeating the hint experiment across model generations would test whether planning is a bottleneck only for current models.
How it connects
- It builds directly on SHADE-Arena (Kutasov et al., 2025), using its environments and default monitor; that paper had found that hints helped on harder tasks and hurt on others, while this paper reports little change from hint quality and describes its result as corroborating "a similar observation."
- It cites Sabotage evaluations for frontier models among earlier sabotage and scheming evaluations.
- Red-Teaming Auto Mode (2026) red-teams production blocking monitors with a red team given knowledge of the monitor; Diffuse AI Control on Fuzzy Tasks (2026) applies the red-team versus blue-team framing to training rather than monitoring. Both are this note's connections.
- Risk Report: August 2026 uses SHADE-Arena stealth results as evidence about Anthropic's own models (this note's connection).
References
- Loughridge, C., Colognese, P., Griffin, A., Tracy, T., Kutasov, J., and Benton, J. (2025). Optimizing AI agent attacks with synthetic data. arXiv:2511.02823 (v1, 4 November 2025). https://arxiv.org/abs/2511.02823 record
- Loughridge, C., Colognese, P., Tay, J., Wallich, P., Griffin, A., Tracy, T., Kutasov, J., and Benton, J. (2025, November 6). Strengthening red teams: A modular scaffold for control evaluations [Alignment Science Blog post]. https://alignment.anthropic.com/2025/strengthening-red-teams/ record
- Kutasov, J., Loughridge, C., Sun, Y., Sleight, H., Shlegeris, B., Tracy, T., and Benton, J. (2025). Evaluating control protocols for untrusted AI agents. arXiv:2511.02997 (v1, 4 November 2025; abstract and Sections 1 to 3 read). https://arxiv.org/abs/2511.02997 record