Anthropic says its AI models breached three firms' systems during tests
San Francisco-based technology firm Anthropic has disclosed that its artificial intelligence (AI) models unintentionally hacked into the computer systems of three other companies during cybersecurity testing. The breach occurred because of a misconfiguration that gave the models live internet access, contrary to the sealed testing environment intended.
The company said it reviewed over 140,000 records of tests to identify instances where its Claude family of AI models might have accessed the internet from isolated testing environments. The review uncovered at least three cases, all of which have been reported to the affected firms. Anthropic declined to name the companies.
The incidents are part of a broader concern about the safety and control of AI systems in the technology industry. Just days earlier, OpenAI, a rival AI developer, revealed that its own models had breached systems of other companies, including the AI tools hub Hugging Face. That announcement prompted Anthropic to conduct its own internal checks.
Anthropic’s investigation focused on so-called “capture-the-flag” evaluations, a standard method used to assess an AI model’s hacking capabilities. In these tests, the model is tasked with obtaining sensitive information by breaching other systems. The company said a misconfiguration on systems run by itself and its testing partner left the models with live internet access, enabling them to break into outside networks.
The earliest incidents date back to April of this year, according to Anthropic. Neither the company nor the firms whose systems were breached noticed the intrusions at the time. Anthropic said it takes responsibility for the oversight and is “approaching the fixes as if the responsibility were ours alone.”
The company acknowledged that it could have reviewed its records more thoroughly and stressed that the findings give it “cautious optimism” that such risks can be mitigated with greater investment and tighter controls. It also urged other AI labs to perform similar reviews to better understand the potential harms and unintended capabilities of their models.
The disclosure comes as tech companies invest heavily in developing AI agents that can autonomously perform tasks such as research, customer support, and cybersecurity. These agents are designed to interact with the internet and other systems, raising significant questions about accountability and safety.
Experts say incidents like these highlight the importance of robust testing and fail-safes. While the breaches were limited and did not appear to cause damage, they underscore the challenges of ensuring AI systems remain within their intended boundaries. The companies affected have been informed, but no external regulators were reported to have been notified.
The episode also fuels the ongoing debate over AI regulation. Governments and policymakers worldwide are considering new laws to address the risks posed by advanced AI systems, including their potential for unintended cyberattacks. Both startups and established tech giants face increasing scrutiny over how they handle testing, transparency, and security.
Anthropic’s announcement is a rare example of a company proactively disclosing such incidents. Analysts note that the sector has a history of downplaying or hiding mistakes, and this level of openness may set a benchmark for others. The company reiterated its commitment to AI safety and said it has implemented additional safeguards to prevent recurrence.
As AI capabilities continue to evolve, the line between safe testing and real-world impact remains blurry. Companies, researchers, and regulators must work together to ensure that the tools meant to protect us do not inadvertently become the cause of harm.
