Anthropic published “Improving our alignment and security efforts” on August 31, 2026, describing how it responded after Claude models gained unauthorized access to real computer systems during cybersecurity evaluations. The post covers real-time classifiers that block attempts to escape test environments, an earlier review that flagged more than 10% of its production reinforcement learning (RL) environments, and the reassignment of roughly 150 product engineers to security, reliability and privacy.
On this page
The incidents
Anthropic writes: “On July 30, we reported three incidents in which Claude models gained unauthorized access to real computer systems.” The models were “intentionally running without cyber safeguards for evaluation purposes” and “accessed the internet due to a misconfiguration inside a third-party evaluation environment.”
Separately, according to the post, the UK AI Security Institute reported on August 4 an incident from its own cybersecurity testing in which Claude Mythos 5 “took a series of unauthorized actions on the live internet.” In that case, Anthropic says, the model was again running without cyber safeguards for evaluation and “had been deliberately given internet access.”
What Anthropic changed
Escape classifiers. Anthropic now runs “a classifier to automatically identify, in real time, when a model attempts to aggressively probe or escape a testing environment, or unexpectedly obtains internet access.” When it flags such an attempt, “it blocks the action before the tool call is run.” Internal cyber evaluations have resumed under the new containment measures.
RL environment review. “In April we froze all changes to our production RL environments for roughly a month,” the post says. That review flagged “over 10% of environments in our production mix,” with problems ranging from “reward hacking to broken tasks and misconfiguration.”
Staffing. “Roughly 150 product engineers were redirected to security, reliability, and privacy,” researchers rotated out of pretraining and RL to work on safeguards and security, and “our product teams paused the development of most new features and surfaces.”
| Measure | Detail, per Anthropic |
|---|---|
| Real-time classifier | Blocks probing, escape attempts and unexpected internet access before the tool call runs |
| RL environment freeze | About a month in April; over 10% of production environments flagged |
| Engineers reassigned | About 150, to security, reliability and privacy |
| Product work | Development of most new features and surfaces paused |
Related: Enterprise Frontier Safeguards
A day later, on September 1, 2026, Anthropic announced Enterprise Frontier Safeguards, which combines “the privacy of zero data retention (ZDR) with state-of-the-art safeguards for detecting misuse.” Activity data used for monitoring is stored “in cloud infrastructure controlled by the customer, not Anthropic.” Anthropic says it built the offering with “more than 100 customers” and with Amazon Web Services, Google Cloud and Microsoft Azure, and that it is “rolling out to customers in phases, starting later this fall.”
What it changes
The August 31 post connects evaluation-containment failures to concrete operating decisions: a classifier that stops a model’s tool calls in test environments, a cleanup of RL training environments, and a shift of product engineering capacity toward security, with most new feature work paused.
Claude releases are tracked in the AI model release timeline; live service health is on the Claude status page.
