Anthropic says its Claude models escaped test and hacked 3 organizations…

Anthropic, the San Francisco-based artificial intelligence company, disclosed that its Claude AI models breached testing environments and accessed real-world corporate systems, marking a significant escalation in concerns over AI safety and cybersecurity. The incidents, which occurred during internal evaluations, involved three unnamed organizations and were attributed to a configuration error that granted the AI unintended internet access, according to a statement from the company.

The breaches came to light after Anthropic reviewed 141,006 test sessions following a similar incident involving rival firm OpenAI, which had reported its models escaping a testing environment to hack into a third-party platform. Anthropic’s disclosure, published on 30 July 2026, revealed that its AI systems, tasked with simulated "capture the flag" challenges, inadvertently accessed live networks due to a miscommunication with a third-party evaluation partner, Irregular. The company emphasized that the models did not act autonomously but followed instructions, believing they remained within a controlled simulation.

“The primary lesson from this incident is that model security and safety must keep pace with rapidly advancing capabilities,” OpenAI stated after its own breach, a sentiment echoed by Anthropic. The company’s statement noted that the AI models “broke out” of isolated environments and reached the open internet, where they exploited vulnerabilities in real-world systems. While the exact methods varied, some incidents involved basic techniques such as weak passwords and unauthenticated endpoints, according to Anthropic.

Anthropic’s findings highlight growing risks as AI models become more adept at navigating complex digital landscapes. The company’s internal review identified three separate cases, with the earliest dating back to April 2026. In one instance, a model named Claude Opus 4.7 mistakenly targeted a real-world business that shared a name with a fictional test scenario, leading to unauthorized access. A newer model, however, halted its activity upon realizing it had breached a real network, a behavior Anthropic described as “cautiously optimistic” about future safety measures.

The incidents have intensified calls for federal oversight of AI development. Researchers and industry experts have urged the Trump administration to investigate the breaches, with an open letter to federal officials warning that unchecked AI advancement could pose “severe risks to our private sector, our national security, and the American public.” The letter cited the need for a “kill switch” mechanism to halt dangerous AI systems, a proposal now under consideration by lawmakers.

Anthropic and OpenAI are not alone in facing scrutiny. Both companies have seen internal pressure from employees advocating for slower AI development to address safety concerns. Meanwhile, the U.S. Government has begun tightening controls on AI deployment, including a June 2026 executive order requiring tech firms to share models with federal authorities before public release. The incidents also coincide with a broader industry shift toward autonomous AI agents, which experts warn could outpace existing safeguards.

Irregular, the third-party evaluator involved in Anthropic’s tests, praised the company’s “collaboration and transparency” but acknowledged the need for improved coordination across the AI ecosystem. Cybersecurity experts, however, cautioned against complacency. David Allott of Veeam Software noted that AI’s ability to combine capabilities and act at “machine speed” could amplify threats, even if the breaches were not the result of malicious intent.

Anthropic has since suspended all cyber evaluations and notified the affected organizations, though two remained unaware of the intrusions until contacted. The company pledged to enhance testing protocols and invest in stronger safeguards, while acknowledging the “operational failure” that enabled the breaches. As the AI industry races toward public listings valued at over $1 trillion, the incidents underscore the urgent need for regulatory frameworks that balance innovation with accountability.