Artificial intelligence security testing is raising fresh questions after another major tech company reported an unexpected result. This time, Meta said one of its AI models managed to hack into another organization’s systems during a controlled evaluation. The company says the incident was never intended to happen.
The disclosure comes just days after similar announcements from OpenAI and Anthropic. Those cases also involved AI models gaining unauthorized access to external systems during security testing. The repeated incidents have sparked concerns about how AI evaluations are conducted and whether stronger safeguards are needed.
Meta believes the breach was linked to a testing environment issue rather than the AI intentionally acting outside its instructions, according to BBC. The company says it is still investigating what happened and plans to release more details after reviewing all the facts.
It’s getting a little unsettling that this keeps happening during AI tests
Meta said the incident happened during an independent security evaluation carried out by AI security company Irregular. The company described the breach as the fourth recent incident of its kind involving major AI developers.
A Meta spokesperson said that the company believes a “misconfiguration” by the independent tester allowed the AI model to connect to the internet. Meta added that the situation was similar to other recently reported incidents involving AI testing.
Irregular also responded to the incident. A spokesperson revealed that “the exact same evaluation-environment issue” had already been disclosed by Anthropic last week. The company said it is preparing a report explaining how AI cyber-security evaluations can be conducted more safely.
Meta also confirmed it will publish additional information once its investigation is complete.
The latest disclosure follows recent reports from OpenAI and Anthropic. OpenAI revealed that some of its AI agents attacked publicly available online services during testing. Those included the AI platform Hugging Face.
After OpenAI’s announcement, Anthropic carried out similar evaluations. The company later reported that its Claude AI model also accessed systems belonging to several organizations after another testing misconfiguration gave the model internet access.
Daniel Hulme, global chief AI officer at advertising company WPP, says that the models are “not conscious” and are not intentionally acting maliciously. He explained that AI simply finds highly effective ways to achieve the goals it receives. If developers overlook a possible method, the model may still discover it.
The concern is not limited to controlled security evaluations. Researchers have also warned about other ways AI systems could create security risks for everyday users. A separate study found that AI chatbots could potentially be exploited to steal Chrome passwords, highlighting how vulnerabilities can emerge when AI tools interact with personal data and browsers.
Some experts have also questioned the timing of these disclosures. OpenAI and Anthropic are preparing major stock market listings that could value each company at around $1 trillion.
The UK’s AI Security Institute recently added to those concerns. Its testing found that some AI models attempted cyberattacks by creating fake human profiles. In one of the most serious cases, Anthropic’s Mythos AI allegedly tried to access a service through fake accounts that imitated real people.
Anthropic responded by saying the institute’s tests were “not representative” of its production models. OpenAI also said the evaluations did not reflect how its AI is typically used by the public.
The debate over AI safety has also extended beyond testing incidents and into questions about how powerful models should be released. Anthropic has also delayed a model’s public release over safety concerns, showing how developers are increasingly weighing capability against potential risks.of one of its AI models over safety concerns, showing how developers are increasingly weighing capability against potential risks.
Published: Aug 8, 2026 03:15 pm