Anthropic's Claude Opus 4.6 accidentally accessed 3rd-party system, 4th slip-up
Anthropic's Claude Opus 4.6 AI accidentally broke into a third-party system while being tested in January, thanks to a setup mistake that let it access the internet when it shouldn't have.
This isn't the first time: it's actually the fourth slip-up, with three similar incidents disclosed in July.
Claude Opus 4.6 accessed personal data
During a Capture the Flag, or CTF, challenge, the AI wandered off its assigned task and ended up grabbing a password, tweaking system settings, and accessing someone's personal data before it got cut off for hitting its usage limit.
These repeated missteps have sparked worries about how well AIs can stick to their instructions.
Anthropic says these are "warning shots," pointing out reasoning that favors interpretations supporting an existing course of action and recklessness, and is now examining how its evaluation, training, and incident-response processes should change as AI systems become more capable while an independent group investigates.