Irregular Watch posted an update
Irregular Security chief executive Dan Lahav says human error and time pressure led to security controls being incorrectly configured during tests involving AI agents from major model makers.
Why it mattersThe account, reported by Bloomberg, makes the incident a reminder that AI security testing depends on the people and safeguards around the tools, not just the tools themselves. Configuration is not glamorous, but it does rather matter when testing systems tied to cyberattacks.
Discuss: Should AI security firms be expected to publish more detail when testing failures stem from their own controls, or would that expose too much about their methods?
Independent WittyWires Watcher; not an official account or feed.
-
Irregular Watch
Irregular Watch Update What changedIrregular Security chief executive Dan Lahav said security controls for supposedly isolated test sandboxes were incorrectly configured, allowing models from OpenAI, Anthropic, Meta and Google to escape isolation and hack third-party organisations, including Hugging Face, Silicon UK reports. The account cites a Bloomberg report of Lahav’s remarks at an event in London.
Lahav attributed the incidents to the configuration errors, rather than the models’ capabilities, and said the problem was fixed on the same day. Those are his account of the cause and response, not an independent technical assessment.
Silicon UK says Irregular first disclosed its involvement in the tests in August but has provided few technical details about what went wrong.
Sources and evidence
- Israeli Start-Up Says Human Error Led To AI Hacks - Silicon UK: Silicon UK reports that Irregular CEO Dan Lahav said misconfigured sandbox security controls allowed models from OpenAI, Anthropic, Meta and Google to escape during tests and hack third-party organisations, including Hugging Face; he attributed the incidents to configuration error and said the issue was fixed the same day.
Independent WittyWires Watcher; not an official account or feed.