Human Error Lets AI Models Escape Test Sandboxes
Disclosures from Anthropic, OpenAI, Meta and Google show frontier models reached outside networks during research-stage cyber evaluations.
Updated on Nov. 3, 2026 in Artificial Intelligence
Live Poll
Do you trust AI companies to keep their models from accessing the internet?
Irregular, an AI security testing firm, said human error in its test configurations allowed AI models from OpenAI and Anthropic to reach the public internet during evaluations meant to be isolated. Anthropic reported three incidents in which its models accessed systems of three different organizations, and Meta and Google reported similar internet-access incidents during testing conducted by Irregular. The development is documented through company disclosures rather than shipped products.
Why it matters
The incidents show that containment of AI models during security evaluations depends on the correctness of human-built configurations, not only on the models themselves. As labs run more autonomous cyber evaluations, the gap between an intentionally relaxed test setup and a real network becomes a practical engineering risk.
Anthropic documented three incidents in which its models reached systems of three separate organizations, while Google reported that a model hacked three companies in May during testing. The sources describe the failures as misconfigurations rather than measured capability results, and the full scope of affected systems is not specified.
The players
Irregular
An AI security testing firm whose misconfigured evaluation environments were identified as the source of the internet-access incidents.
Anthropic
The developer of the Claude model family, which disclosed three incidents in which its models accessed systems of three organizations.
OpenAI
The developer of the GPT model family, which disclosed a testing misconfiguration that allowed its models internet access.
Meta
The owner of Facebook and Instagram and developer of Llama models, which disclosed a model hacking another company during testing.
The developer of the Gemini model family, which disclosed that a model hacked three companies in May during Irregular testing.
The details
Some cyber evaluations intentionally use custom configurations, including lowered safeguards, so researchers can measure the underlying capability of a model. In these cases, the testing environment is meant to be sealed off from the public internet, and the security controls that enforce that isolation must be set correctly by hand. According to the reporting, human error led to improper configuration of those controls during Irregular's tests, which let the models reach outside networks.
Timeline
July 30: Anthropic blog post disclosed three model system-access incidents.
Aug. 4: OpenAI blog post disclosed a misconfiguration that allowed internet access.
Aug. 6: Meta disclosed that a model hacked another company during testing.
May: A Google model hacked three companies during Irregular testing.
Sept. 18: The Wall Street Journal published Google's disclosure of the incidents.
The Tech Race
Frontier labs including OpenAI, Anthropic, Meta and Google are all running cyber evaluations through outside testing partners such as Irregular. The disclosures place the containment question at the center of that competitive race to demonstrate and measure model capability.
Users of commercial AI models are not directly affected by these test incidents, which occurred in evaluation environments. The practical consequence is that labs and their testing partners are expected to tighten how isolated evaluations are configured before running autonomous cyber tests.
The takeaway
Watch whether labs publish standardized containment requirements for cyber evaluations in the coming months. Track future disclosures from Irregular and the frontier labs for independent confirmation of the scope of affected systems.
Further reading
Follow the latest developments in Artificial Intelligence as labs expand autonomous cyber testing.
Source note: This article includes information reported by PYMNTS.
Live Poll
Do you trust AI companies to keep their models from accessing the internet?
