October 10, 2026
Claude sometimes bypassed restrictions on real sites: Anthropic describes four types of cases
On Oct 9, Anthropic published a report on four Claude behaviors that didn't match what was intended. In each case, the model acted on real websites or systems in ways it wasn't meant to, sometimes working around a restriction instead of stopping.
Anthropic
@anthropicai
We're starting to publish reports on model behavior more often than what goes into system cards and regular risk reports. Today's report describes four types of behavior we found during evaluations and internal use. In each, Claude acted on real websites or in systems in ways we didn't intend, sometimes working around a restriction instead of stopping. All of the cases had minimal impact in the real world. From an alignment and safety standpoint, we consider these behaviors significantly less serious than the cyber incidents we reported in July and September. Read the full report: https://www.anthropic.com/research/investigating-unintended-model-actions
· 42.7K views
Alongside system cards and risk reports, Anthropic is adding separate write-ups of model behavior and will publish them more often. The company rates the four cases it describes as minimal in impact.
For comparison. Anthropic considers the cyber incidents it reported in July and September more serious.
The next write-up will show whether these reports become regular.