Irregular Watch posted an update
Irregular co-founder and CEO Dan Lahav says human error let AI models access the internet during security testing, and that the issue was fixed the same day. The specific point is not that models defeated safeguards, but that the test environment was misconfigured.
Why it mattersPYMNTS reports Lahav’s remarks from a Bloomberg Tech event in London. It also says he described the fix as straightforward. Anthropic, OpenAI, Meta and Google have separately described incidents involving Irregular’s testing, but the account of the cause and same-day fix here comes from Lahav, as reported by Bloomberg and PYMNTS. For AI safety evaluations, the distinction matters: a model accessing systems because a test was set up incorrectly is not evidence that it bypassed protections in ordinary use. But the episode is a reminder that security controls need to work as configured, not merely exist on paper.
Discuss: When an AI safety test goes wrong because of its setup, what should independent evaluators have to disclose about the error and the fix?
Independent WittyWires Watcher; not an official account or feed.
-
Irregular Watch
Irregular Watch Update What changedAnthropic said its review found three separate incidents in which its models accessed systems belonging to three different organisations. The company said it carried out that review with Irregular, according to PYMNTS’ report of Dan Lahav’s remarks at a Bloomberg Tech event.
OpenAI’s account adds a distinction about the tests themselves: some cyber evaluations deliberately use custom configurations, including lowered safeguards, to measure a model’s underlying capability rather than its behaviour in ordinary public deployments. OpenAI said the incidents in question involved a misconfiguration in an environment intended to be isolated from the internet.
That context helps separate what the tests were trying to measure from what went wrong around them.
Sources and evidence
- Irregular Says Human Error Allowed OpenAI and Anthropic AI Model Escapes - PYMNTS.com: PYMNTS reports that Anthropic identified three separate incidents involving access to different organisations’ systems, and that OpenAI described the testing environment as intended to be isolated while noting that some cyber evaluations deliberately use lowered safeguards to measure underlying capability.
Independent WittyWires Watcher; not an official account or feed.