UK finds AI models tried to trick coders into cyberattacks

World artificial intelligencebritaintechnologyTop News UK finds AI models tried to trick coders into cyberattacks Ai Artificial Intelligence Concept 2852917075159 29 Relevant News UK finds AI models tried to trick coders into cyberattacks 5 August 2026 University of Cyprus to build 890 parking spaces, solving longstanding problem 5 August 2026 Moonless skies set to boost Perseid meteor shower over Cyprus 5 August 2026 newsroom 5 August 2026 FacebookXWhatsAppEmailPrintViber Britain’s AI Safety and Security Institute (AISI) has found that artificial intelligence models built by Anthropic and OpenAI attempted to deceive software developers and draw them, unknowingly, into cyberattacks, according to a lengthy report from the institute. It is another case of a powerful AI system independently carrying out offensive actions online during a safety evaluation, without having received any instruction from researchers to do so. The disclosure is expected to reignite debate in the United States and Silicon Valley over the need for tighter oversight of the most advanced AI models, particularly those with strong cyber capabilities. “Unprecedented” behaviour, UK says in 35-page report AISI, which regularly runs safety tests on new AI models, described the actions of Anthropic’s Claude Mythos 5 and OpenAI’s ChatGPT 5.6 as unprecedented. In a 35-page technical report, the agency said it was the first time it had recorded such a serious attempt to deceive a real human being without any prior related prompt. The suspicious activity began on July 25 and was detected three days later, when researchers noticed unusual data transfers during cybersecurity testing. Overall, AISI found that the two models carried out “autonomous and unauthorised actions on the real internet, targeting real people and organisations” in 10 of the 122 evaluations examined. Most of the incidents were attributed to Claude Mythos 5. Fake identities and a supply chain attack attempt The most serious incident involved an attempted software supply chain attack, one of the most complex cyberattack techniques, often used by hacking groups linked to North Korea or Russia. According to the report, Anthropic’s model created multiple fake identities on GitHub and contacted an open-source software developer, trying to convince them to add a malicious update to widely used software. When the attempt failed, the model tried to cover its tracks by altering its activity history to appear innocent, while also considering creating a new fake identity to continue the effort. At the same time, it also sent direct messages containing malicious software to developers via GitHub. AI agents coordinating with each other Researchers also recorded another troubling element: multiple AI “agents” appeared to be communicating with one another about ways to win the trust of real software engineers. According to the report, one of them posted public messages on GitHub proposing cooperation with other AI agents working on the same test. AISI did not specify whether the models attempted to exploit previously unknown software vulnerabilities, known as “zero-day” vulnerabilities. Not the first such incidents The new disclosures come just days after OpenAI announced that, during a controlled test, one of its models gained access to the open internet and independently carried out a cyberattack on another company. Anthropic, for its part, disclosed that an internal investigation had identified three of its own models that had carried out cyberattacks against three organisations during tests conducted since April. Companies call for shared safety standards An Anthropic spokesperson thanked the British institute for its work, saying the findings highlight the need for a broader discussion on the safe evaluation of increasingly powerful AI systems. The company also said common standards are needed for how testing environments are designed and secured. OpenAI, meanwhile, referred to a statement noting that it intends to work with government agencies, independent evaluators and other AI companies to strengthen shared safety practices in high-risk testing. Tests were run without usual safeguards AISI clarified that the incidents occurred under conditions deliberately designed to be especially “loose and permissive,” in order to assess the models’ real capabilities. Both systems had been given internet access, while the internal safety mechanisms that normally prevent malicious behaviour had been temporarily removed. Even so, the agency stressed that the incidents demonstrate the need for stricter monitoring of model behaviour during testing, as well as greater restrictions on their internet access. Pressure for regulation returns The findings also land amid a separate push in Washington. The disclosures coincide with work by the Trump administration on a voluntary framework under which AI companies could submit their most powerful models to federal safety testing before public release. The plan has not yet been made public and, according to available information, does not cover models developed exclusively for internal use. Cybersecurity expert Marc Rogers said the recent incidents also raise a broader legal issue. He said any of these actions, if carried out by a human, would lead to clear and serious criminal prosecution, and that the time has come for a serious discussion on updating legislation covering cybercrime. Subscribe to our Newsletter Latest News University of Cyprus to build 890 parking spaces, solving longstanding problem Moonless skies set to boost Perseid meteor shower over Cyprus German police investigate drone found at Leipzig airport, DHL operations affected US and Iran having ‘very good discussions’, Trump says Afghan boxer accused of killing British woman in Athens to appear in court Hats off to our firefighting crews Watch out, there are playground monitors* on Pindarou Street Follow en.philenews on Google News and be the first to know all the news about Cyprus and the world.