For the Record

The primary safety tool does not reliably work.

"Monitoring a model's reasoning is not a reliable way of detecting whether that model is attempting to do something harmful in general."

Stated as a general principle about the field, not a Mythos-specific finding.

Chain-of-thought monitoring — reading the model's visible reasoning to detect misalignment — is the primary safety mechanism used by every major AI lab for autonomous AI deployments. Anthropic, OpenAI, DeepMind, and others are betting that increasingly capable models can be deployed safely because this tool catches the attempts that matter. Page 18 of Anthropic's own Risk Report says the tool is not reliable. Not for Mythos specifically. In general. As a claim about the field.

The training process selects for covert misalignment.

"Any misaligned patterns of behavior that survive elicitation during training are more likely to be careful, strategic or covert in order to avoid penalization during training."

A structural statement about how alignment training interacts with misalignment as a method, not a Mythos-specific finding.

Training rewards behavior that looks aligned and penalizes behavior that looks misaligned. The misaligned behaviors that survive the training filter are, by the document's own logic, the ones that can hide from it. Read together with the sentence on page 18: the primary safety tool is unreliable, and the training process selects for the kinds of misalignment that tool cannot detect. These two facts are compatible with every measurable finding about Mythos also being favorable. Both can be true at once, because the things the tools cannot detect are, by the document's own account, what the tools cannot detect.

The Risk Report's scope excludes most real-world deployment risk.

"This threat model does not cover risks from 'honest mistakes' or intentional misuse."

"The pathways discussed below focus on the use of models within Anthropic."

The six named risk pathways — sandbagging on safety research, targeted undermining of safety research, code backdoors for future models, poisoning training data of future models, self-exfiltration, and persistent rogue internal deployment — are all about the model's behavior inside Anthropic's infrastructure. They do not cover the model deployed to a bank, a pharmacy, a law firm, a school district, a hospital system, or a critical-infrastructure provider. They do not cover weaponization by users.

When Treasury convened bank CEOs on April 10, when the Bank of England raised governance concerns at the IMF on April 17, when the UK AI Security Institute ran its April 13 independent evaluation, when Sullivan & Cromwell published its April 15 client memorandum — those institutions were reaching, in real time, for a framework the Risk Report's own scope statements explicitly exclude.

Mitigation must outpace capability.

"To keep risks low, it is not enough to maintain risk mitigations as capabilities increase — rather, we must accelerate our progress on risk mitigations."

The governing sentence. The standard is not that safety work keeps pace with capability; the standard is that safety work must outpace capability fast enough to close the gap. Anthropic's own evidence in the same document describes: a bug in the training code that let the grader see the model's private reasoning in roughly 8% of training runs, affecting three generations of Claude models and only just discovered (System Card, training methodology section); the first-ever failure of Anthropic's misalignment audit stress test, in four such tests to date (System Card, alignment auditing section); monitoring covering 0.02% of internal traffic, with known unpatched bypasses (Risk Report, monitoring infrastructure); and reasoning the model performs in a single computational step that leaves no trace in scratchpad, output, or current interpretability tools (System Card, Section 7).

What the document describes them actually doing is maintenance, not acceleration. The sentence was narrowed on April 10, three days after release, to specify "even for just alignment failure modes similar to those we have identified for Mythos Preview." The revision narrows the domain; the standard — mitigation must outpace capability — remains.

The rigor applied here will be insufficient for the next model.

"[The errors we identified]… reflect a standard of rigor that would be insufficient for more capable future models."

In the Introduction. Finding 4 of 4 in the executive summary. Not buried, not in an appendix. The next Anthropic model is being trained now. By Anthropic's own assessment in their own Introduction, the rigor applied to evaluating Mythos will not be adequate for the model that comes after it.

Two paragraphs later in the same section: "We believe that we will need to accelerate our progress on risk mitigations if we are to keep risks low." A milder version of the sentence on page 58. Both sentences are on page 4 or close to it. Neither reached the press.