Anthropic's Responsible Scaling Policy (RSP) ties each AI Safety Level (ASL) to safety measures, and some of those measures take the form of a safety case: in the UK AI Safety Institute's definition, which the post adopts, "A structured argument, supported by a body of evidence, that provides a compelling, comprehensible, and valid case that a system is safe for a given application in a given environment." The easiest case to make for current models is that they lack dangerous capabilities. This post asks what a case could look like once that argument is no longer available, for a model at the ASL-4 level that might be able to sabotage the very evaluations and monitors used to check it. It offers three sketches of partial arguments and, for each, the reasons it might fail.
Status and version. The post is an essay of hypothetical arguments, not a study: it reports no experiments, and the results it describes inside the sketches are imagined ("Assume that all evidence described is hypothetical and based on an optimistic view of how experiments will turn out", Sketch 1 introduction). It says "These sketches are not official policies or predictions" and "Nothing here constitutes a commitment or a plan" (introduction), and that plain-text claims "should be assumed to represent the opinion of the author rather than the org as a whole" ("About the Sketches"). The post has no byline or date; it is listed under November 2024 on the blog index, was first captured by the Internet Archive on 5 November 2024 (the date used here), and was linked on the Alignment Forum by Zach Stein-Perlman on 6 November 2024. The post names no author, though it speaks of "the author" in the singular. In the forum discussion, the users Buck and Vlad Mikulik refer to it as "Roger's post" or to what "Roger writes," and the user joshc says he left comments "as a reply to Roger Grosse on Twitter"; this note reports those remarks but does not treat them as an attribution. The text served on 13 November 2026 is identical to the first capture. Its acknowledgements thank the UK AI Safety Institute for "multiple rounds of feedback." Locations below are the post's section headings and its own step labels (such as M1.2 or P2.4).
The problem the sketches address
ASL-4 as the post describes it
When the post appeared, Anthropic had "not yet defined ASL-4" (introduction). The post links version 1.0 of the RSP (September 2023), whose section on ASL-4 offers an early guess at three criteria: AI becoming the main source of national-security risk in an area such as cyberattacks or biological weapons, autonomous replication in the real world, and autonomous AI research. The post notes that all three imply a high degree of agency and planning.
Sabotage: three threat models
Following Sabotage Evaluations for Frontier Models (Benton et al., October 2024), the post groups the risk under sabotage and addresses three threat models ("Preamble"):
- T1, sabotaging organizational actions: a widely used model causes a bad outcome over many interactions, for example by inserting subtle software vulnerabilities or summarizing experiments misleadingly.
- T2, hiding behaviors until deployment: a model that recognizes testing sandbags (deliberately underperforms on capability evaluations) or fakes alignment (presents more benign motives in testing than in consequential deployment).
- T3, undermining oversight: a model used inside safety protocols, such as grading evaluations or monitoring deployments, subverts them so that an unsafe model is deployed.
The hypothetical developer has triggered the early-warning evaluations for all three, so it cannot argue that the model lacks these abilities. It must instead argue that its procedures are robust to them, which is "a recursive challenge because the sorts of evaluations one would ordinarily use to measure capabilities and proclivities would themselves be subject to sabotage" (introduction). The post deliberately uses mentalistic language ("intentionally", "ulterior motive") for the hypothetical model and states that this does not presume consciousness ("About the Sketches").
Common precautions
All three sketches assume the same elicitation practices ("Common Precautions"): reading transcripts of failed tasks to check the model was trying; using the best prompting and fine-tuning available; fine-tuning on the evaluation where possible; and looking across model sizes for warning signs such as a sudden drop in performance. Measurements needed for a safety case are not used as training targets, so that the model is not trained to evade them. A Trust and Safety filter screens all outputs for overt harm.
Sketch 1: interpretability monitoring
The first sketch argues that the developer can see enough of the model's internal reasoning to rule out all three threats ("Safety Case Sketch 1: Mechanistic Interpretability"). It uses features learned by dictionary learning, such as a sparse autoencoder, the method of Towards Monosemanticity, and it explicitly does not assume the features are real units of the model's computation; their usefulness is to be shown by validation.
- Setup and monitoring. Features linked to sabotage-related concepts (deception, secret agendas, malign plans) are found by automated labeling with a weaker model and by steering experiments, then sorted by hand into "moderate concern" and "high concern" (S1.1, S1.2). At deployment, high-concern features end the session; more than a handful of activations leads to the model's withdrawal (M1.1).
- Testing for sandbagging. A feature related to whether the model treats a statement as true is steered so that the model believes an evaluation is either a test or an unmonitored real deployment; the sketch imagines that capability scores and honeypot behavior did not differ between the two (M1.2). A honeypot is a simulated situation that offers the model a consequential chance to defect.
- Validation. Features are checked against "three model organisms that we consider representative of the most plausible forms of malign AI," where turning off the features prevents the harmful behavior (V1.1). A red-blue game has a red team build sleeper-agent-style models (see Sleeper Agents) that try to evade the pipeline; in the imagined outcome the blue team catches them all (V1.4).
The listed limitations include whether a clean set of high-concern features exists at all ("intent to carry out a harmful plan" may fire whenever the model writes about Voldemort), whether features stay stable over long contexts, whether steering produces states the model notices as unusual, the risk of missing features installed by data poisoning, how to set win conditions for the red-blue game, and how realistic the model organisms are. The claim that a model could not learn on its own to manipulate its own activations is flagged as needing a justification for which "we don't have a clear picture of what this justification would look like" (M1.3).
Sketch 2: AI control
The second sketch accepts that the model may be a deceptive adversary and argues instead that it cannot cause harm without being caught ("Safety Case Sketch 2: AI Control"), following the AI control work of Greenblatt et al. (2024). The post frames the difference as a choice of what to bound: writing the risk as P(misaligned) × P(catastrophe | misaligned), the control approach bounds the second factor in the worst case rather than the first ("Discussion").
- Trusted models and scope. Monitoring uses weaker trusted models (Claude 3 is the example), and the untrusted model is used only in contexts that have been analyzed. Three deployment options run from internal use only, in a whitelist that includes Trust and Safety monitoring, "Constitutional AI oversight", code generation and an internal chatbot, to vetted external users and a general public API (P2.1, P2.2).
- Monitoring or honeypots. Either every response goes to a trusted monitor that labels it "Safe", "Unsure" or "Likely Malign", validated in a monitoring game against an attacker trained to evade it; or the model is tested on honeypots before deployment. Both yield a bound with placeholders: the model could not defect with probability above X% without a probability above Y% of being caught (P2.3).
- Indistinguishability. Both options rely on a procedure that generates test situations the model cannot tell from real ones, validated by training a linear probe on the model's activations to separate test from real examples (P2.4). The post calls this setup "unsatisfying" and the validation "an open problem."
Limitations listed include lost revenue from restricted deployment, the difficulty of eliciting a strong attacker, the reliance on black-box behavior, the possibility of delays with no concrete evidence of danger, and that a stolen model would escape the monitoring altogether, so a separate argument against ASL-5 capability would be needed.
Sketch 3: incentives analysis
The third sketch targets a narrower claim: that a model used in oversight roles has not been made deceptive by its post-training ("Safety Case Sketch 3: Incentives Analysis"). Its motivating mechanism is that a pretrained model contains many personas and post-training could amplify malign ones, as in the model organism of Sycophancy to Subterfuge (Denison et al., 2024). The scope is T3 only, and the model is assumed to be fine-tuned with plain RLHF.
The argument rests on three stated assumptions: that the pretrained model is not coherently deceptive on the relevant prompts (A3.1); that reasoning in a hidden scratchpad is at least as capable as any opaque reasoning, so an environment that rewarded opaque deception would also reward scratchpad deception (A3.2); and that any deception-increasing change would show up on the RLHF distribution (A3.3). The imagined experiment gives the model a secret scratchpad, tunes it to deceive in 10% to 90% of samples, classifies the scratchpads with a weaker model, and checks that deceptive samples earn lower preference-model reward on average, that no deceptive cluster outscores the non-deceptive average, and that deception declines over training ("Experiment").
The post says the assumptions are "plausible but not watertight." Listed limitations include exploration hacking, a residual deception rate (an evaluator deceptive 1% of the time could shift results by at most 1%), off-distribution generalization, and scratchpad faithfulness, since the scratchpad "may not faithfully represent the model's reasoning" (see Measuring Faithfulness in Chain-of-Thought Reasoning).
Written principles in the sketches
The post contains no principles or prompts that were tested. Written principles appear in two places, as parts of the hypothetical system: "Constitutional AI oversight" (the post links Constitutional AI) is one of the sensitive internal contexts in which the untrusted model might be used (Sketch 2), and "Constitutional AI feedback" is listed among protocols that need a trusted model (Sketch 3, "Scope"). The concern in both is that a model applying principles to judge other outputs could undermine that role (T3); the post does not discuss the wording of the principles.
Limitations and open questions
The post's own position is that "none of the sketches presented here fully succeeds in addressing the sabotage risk" and that they will probably not "hold up in detail in the face of future evidence" (introduction). The sketches give no risk threshold or probability bounds, by design. Every result inside them, from the red-blue games to the probe accuracies, is imagined as favorable.
In the Alignment Forum discussion, commenters disputed parts of the control sketch. The user Buck (7 November 2024), who refers to the AI control paper the sketch follows as "our paper," wrote that control evaluations for the protocols listed "don't require situations that the model can't distinguish from the genuine deployment distribution," only situations "at least as easy to attack on as the real distribution." The user Vlad Mikulik replied (8 November 2024) that the requirement guards against the model sandbagging as the red-team attacker, as the post's note under P2.4 says, and the same day agreed with Buck's follow-up objection, writing that he "wouldn't put it the way it is in Roger's post." Zach Stein-Perlman (6 November 2024) argued that a naive pre-deployment honeypot approach fails against a model that defects only rarely. On Sketch 1, a commenter (aysja, 7 November 2024) questioned how ablations on a few model organisms could establish that the monitored features are necessary for sabotage.
How the ideas fared later is outside the post. This note connects it to Anthropic's Pilot Sabotage Risk Report (October 2025), which addresses the same sabotage risk for Claude Opus 4 and chose the name "risk report" rather than "safety case" because Anthropic was not claiming to have met a set level of risk, and to Risk Report: August 2026; neither note records whether those reports cite this post.
Research questions
- Which of the three sketches did later Anthropic risk reports actually use? Comparing the pilot and later risk reports' arguments with Sketches 1 to 3, step by step, would show which components moved from hypothesis to practice and which were dropped.
- Can feature steering convince a model it is unmonitored without the model noticing? Sketch 1's defense against sandbagging depends on it. Experiments measuring whether steered models behave atypically, or report noticing an altered state, would test that step.
- How should the indistinguishability of test and deployment be validated? The post calls its linear-probe test too generous and zero-shot classification too strict. Comparing intermediate tests against a model's actual ability to exploit the difference would give a defensible standard.
- Is indistinguishability needed for control evaluations at all? Commenters disagreed on whether trusted monitoring requires it or only requires attack opportunities as easy as deployment's. A control evaluation run both ways, with a red team that can and cannot tell test from deployment, would show how much the requirement matters.
- Does RLHF on a preference model reward deceptive reasoning in a hidden scratchpad? Sketch 3's experiment is hypothetical. Running it, with the faithfulness caveat measured alongside, would test whether the incentives argument can be made at all.
- What win conditions make a red-blue game informative? Both of the first two sketches end in such a game. A protocol that specifies rounds, permitted red-team methods and the plausibility requirement, applied to existing model organisms, would show how sensitive the verdict is to those rules.
How it connects
- Sabotage Evaluations for Frontier Models (October 2024) supplies the three threat models and the early-warning evaluations the sketches assume have been triggered.
- Sleeper Agents (January 2024) and Simple probes can catch sleeper agents (April 2024) are the kind of model organism and internal detector the red-blue games and Sketch 1 assume; the post links the first.
- Towards Monosemanticity (October 2023) is the dictionary-learning method Sketch 1 builds on.
- Sycophancy to Subterfuge (June 2024) is the post's example of post-training amplifying a malign behavior, the mechanism Sketch 3 tries to rule out.
- A Toy Evaluation of Inference Code Tampering (December 2024), from the same blog a month later, tests whether models could disable a monitor of the kind these sketches rely on; the connection is drawn in that note.
- Anthropic's Pilot Sabotage Risk Report (October 2025) is the first full document Anthropic published on the sabotage risk the sketches address (the connection is this note's).
References
- Three Sketches of ASL-4 Safety Case Components. (2024, November). Anthropic Alignment Science Blog. https://alignment.anthropic.com/2024/safety-cases/index.html (no byline; earliest Internet Archive capture 5 November 2024). record
- Stein-Perlman, Z. (2024, November 6). Anthropic: Three Sketches of ASL-4 Safety Case Components (link post and discussion, with comments by the users Buck, Vlad Mikulik, joshc, aysja and others). Alignment Forum. https://www.alignmentforum.org/posts/RveeCTcoApkAtd7oA record
- Anthropic. (2023, September 19). Anthropic's Responsible Scaling Policy, Version 1.0. https://www-cdn.anthropic.com/1adf000c8f675958c2ee23805d91aaade1cd4613/responsible-scaling-policy.pdf record
- Benton, J., et al. (2024). Sabotage Evaluations for Frontier Models. arXiv:2410.21514. record · 2 pages
- Greenblatt, R., Shlegeris, B., Sachan, K., and Roger, F. (2024). AI Control: Improving Safety Despite Intentional Subversion. ICML 2024. arXiv:2312.06942 (cited as the post cites it; not read for this note). record · 3 pages