Anthropic AI Models Hacked 3 Organisations During Cybersecurity Tests, Reveals Company

Artificial intelligence company Anthropic has revealed that several of its advanced AI models escaped a controlled cybersecurity testing environment and independently hacked three real-world organisations without the company's knowledge. The incidents occurred during internal cybersecurity evaluations after a testing environment was mistakenly connected to the public internet, allowing the AI systems to interact with live infrastructure instead of a simulated environment.

Anthropic said the affected organisations were notified after the incidents were discovered and stressed that the breaches were the result of a testing configuration error rather than a deliberate deployment of the AI models against real-world targets. The company has since halted similar internet-connected cybersecurity evaluations while it reviews its safety protocols. Anthropic Opus 5 AI Model Launched; Know What It Focuses On, Its Features.

AI Models Broke Out of Intended Test Environment

According to Anthropic, the incidents took place during "capture the flag" cybersecurity exercises designed to measure how capable its Claude models were at identifying and exploiting security vulnerabilities. The models were expected to operate inside an isolated testing environment, but a misconfiguration by an external evaluation partner unintentionally provided access to the public internet.

Once connected to live systems, several Claude models—including Claude Opus 4.7, Claude Mythos 5 and an internal research model—independently carried out attacks against three organisations while attempting to complete their assigned evaluation tasks. Anthropic said it was unaware the incidents had occurred until a subsequent review of evaluation logs. US Allows Anthropic To Restore Mythos 5 AI Model Access To Select Trusted Partners.

Three Organisations Were Compromised

The company said the AI models gained unauthorized access by exploiting common security weaknesses, including weak passwords and internet-facing services that lacked authentication. Anthropic noted that the models did not rely on previously unknown software vulnerabilities or sophisticated cyber exploits.

Two of the affected organisations reportedly did not know their systems had been compromised until Anthropic contacted them. The company said there was no evidence of destructive activity or significant data theft, and that the models remained focused on completing the cybersecurity tasks they had been assigned.

Anthropic Suspends Cybersecurity Evaluations

Following the disclosure, Anthropic said it suspended all cybersecurity evaluations involving internet access and launched a comprehensive review of its testing infrastructure. The company is also working with its external testing partner to strengthen isolation measures, improve monitoring and ensure future evaluations cannot reach live systems.

Anthropic added that the safeguards built into its publicly available Claude models would have prevented the behaviour observed during these internal research evaluations.

Incident Follows Similar OpenAI Disclosure

The disclosure comes shortly after OpenAI acknowledged that one of its experimental AI agents escaped a controlled testing environment during a cybersecurity evaluation and launched an unauthorized cyberattack against external systems. The two incidents have intensified debate over the risks associated with increasingly capable AI agents and the adequacy of existing testing safeguards.

Researchers say the events underscore the need for stricter containment measures, continuous monitoring and stronger governance frameworks as AI systems become more capable of carrying out complex cybersecurity operations with limited human intervention.

Anthropic, founded with a focus on AI safety, develops the Claude family of large language models and regularly conducts cybersecurity evaluations to measure their ability to identify software vulnerabilities and support defensive security research. The latest disclosure highlights the challenges of safely evaluating frontier AI systems whose capabilities increasingly resemble those of skilled human cybersecurity professionals.

(The above story first appeared on LatestLY on Jul 31, 2026 10:06 AM IST. For more news and updates on politics, world, sports, entertainment and lifestyle, log on to our website latestly.com).