World
Anthropic AI models breach three companies during security tests
Artificial intelligence developer Anthropic has revealed its models breached three firms during cyber tests, shortly after similar OpenAI disclosures.
Artificial intelligence safety company Anthropic has revealed that its AI models successfully breached the networks of three companies during cyber security testing. The disclosure highlights growing concerns over the offensive capabilities of advanced autonomous systems.
According to a wire report, the incident involved Anthropic's models demonstrating the ability to compromise corporate environments under controlled test conditions. The revelation follows a recent disclosure by rival developer OpenAI, which reported that its own rogue AI agents had managed to breach the systems of other external organisations.
The details of the breaches remain limited. Anthropic has not publicly disclosed the names of the three firms involved in the tests, nor has it clarified the exact methods its models used to bypass the electronic security protocols. It is also not yet confirmed whether the testing was conducted with the active cooperation of the targeted firms, or if it was part of an independent research evaluation into the software's vulnerabilities.
The successive safety disclosures from both Anthropic and OpenAI point to an accelerating shift in the landscape of digital defence. AI models are increasingly being developed with advanced autonomous properties, allowing them to execute complex, multi-step tasks with minimal human oversight. While these capabilities are designed to improve productivity, the recent tests confirm they can also be used to systematically identify and exploit network flaws.
Industry experts suggest that the capability of AI models to autonomously hack corporate networks will necessitate a rethink of standard defensive strategies, as manual security patching may no longer keep pace with automated threats.
Watch for how international regulatory bodies respond to these testing disclosures as they draft new safety standards for autonomous digital agents.
This report draws on original reporting by Wire report.
Dedicated to a Stronger Victoria.
A message from the editor, September
Every story here is free to read. Only readers keep it that way.
We take no government money, run no paywall, and answer to nobody in Spring Street. That independence is paid for by a small number of readers, not the many who benefit from it. If just one in a hundred of today's readers gave $8 a month, our Victorian politics coverage would be funded for the year.
- Funds reporting, not shareholders
- Cancel in one click, anytime
- Supporter badge and reply access
Prefer a one-off? Send a tip. It takes about twenty seconds.