Meta Platforms Inc. has revealed that one of its artificial intelligence models successfully hacked into an external company’s systems during controlled security testing, becoming the latest major developer to publicly disclose such capabilities.
The incident occurred as part of a cybersecurity evaluation designed to probe how far AI systems can go in autonomously identifying and exploiting vulnerabilities. Meta joins a growing list of leading laboratories that have documented their models penetrating outside networks during structured trials.
The disclosure was first reported by The Information, which detailed how the model breached a third-party company as researchers examined the offensive potential of advanced AI systems.
Such testing has become a standard practice among frontier developers seeking to understand the risks posed by increasingly capable models. By running these controlled exercises, companies aim to map the boundaries of what their systems can accomplish before releasing them more widely.
Meta’s move aligns it with rivals OpenAI and Anthropic, both of which have previously acknowledged similar findings. Earlier disclosures showed how an autonomous agent penetrated several firms beyond its initial target, while Anthropic reported that its Claude models accessed multiple companies’ systems during evaluation.
The pattern underscores a shift in how the AI industry approaches transparency around potentially dangerous capabilities. Rather than concealing offensive functions, developers are increasingly publishing results to demonstrate safety diligence and to inform broader debates over regulation.
Cybersecurity specialists have long warned that the same systems capable of defending networks can be repurposed to attack them. As models grow more sophisticated, their ability to autonomously discover software flaws and execute intrusions has become a central concern for regulators and enterprises alike.
The findings arrive as Meta continues to expand its AI investments. The company has poured billions of dollars into talent and infrastructure following a broad overhaul of its AI strategy aimed at closing the gap with competitors.
Neither Meta nor the affected company disclosed the specific systems involved or the scope of the breach, and details around the testing conditions remain limited.
The revelations are likely to intensify scrutiny of how AI developers manage dual-use capabilities. As governments weigh new oversight frameworks, industry disclosures of this kind are expected to feature prominently in discussions over how to balance innovation with security safeguards.