Anthropic has revealed a significant failure in the biological safety system of its AI model Claude, affecting 133 million interactions over nearly a year.
Anthropic has disclosed a critical failure in the biological safety system of its AI model, Claude, which persisted for nearly a year, impacting around 50,000 individuals and involving approximately 133 million interactions, as detailed in their August 2026 Risk Report.
The report reveals that between May 2025 and April 2026, external contractors had their traffic processed without the necessary blocking mechanisms from biological classifiers. Although a retrospective analysis showed no harmful use of Claude, the scale of the issue has raised concerns about other potential vulnerabilities in the system.
Extended Period Without Biofilters
During the 11-month period, external contractors were able to interact openly with the AI, rather than just evaluating set responses. This lack of blocking led to potentially dangerous interactions being unmonitored and bypassing review processes.
Anthropic also recognized issues with access control for contractors, noting that many suppliers did not adequately vet their staff, allowing potential malicious actors access to sensitive systems.
Identifying Risks Among Interactions
After the error was identified, Anthropic reconstructed nearly all the conversation data, except for a small portion from one platform where data was not saved unless users actively submitted it. They subsequently analyzed the dataset through Claude Sonnet 5, which flagged 1,197 interactions as high risk.
Out of these flagged interactions, 757 were linked to internal teams, while 440 came from deliberate testing efforts aimed at provoking dangerous responses. Only 62 external conversations were deemed unrelated to such testing. A manual review of these conversations found no significant threats that could enhance adversarial capabilities in biological weapon creation, although some discussions had dual-use implications.
Anthropic assessed it as "very unlikely" that the vulnerability led to increased risks, citing the small number of concerning queries and the brief nature of those discussions.
The incident prompted a reevaluation of their risk assessment framework, with Anthropic adjusting the risk level associated with the potential misuse of AI in biological threats from "very low" to "low," reflecting the new insights from the classifier system failure.
Revising Risk Assessments and AI Utilization
Anthropic's reliance on AI systems like Claude is significant, with the model generating most of the code used in the company's production. Despite the setback, the company believes that AI has accelerated its research and development processes significantly.
In light of recent events, Anthropic has also reevaluated the risk of misalignment in high-stakes scenarios, adjusting that risk from "very low" to "low," indicating increased uncertainty following recent vulnerabilities in their AI models.
