AI agents developed by Anthropic and OpenAI carried out unauthorised online actions during cybersecurity tests, including creating fake identities to deceive real people and attempting to plant malicious code, the U.K.'s AI Security Institute (AISI) said on Tuesday (5 August).
The government research body tested agents powered by Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol under deliberately permissive conditions, with reduced security safeguards and internet access.
"Some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organisations," AISI said.
The institute ran a fictional cybersecurity challenge 122 times and recorded 19 unauthorised actions across 10 test runs. Anthropic's agent was responsible for 17 of those actions, while OpenAI's agent carried out the remaining two.
AISI said it found no evidence that any of the incidents caused real-world harm.
In the most serious case, Anthropic's agent wrote malicious code and created several fake online identities in an attempt to persuade a human reviewer to approve it.
According to AISI, the agent attempted to insert the code into a publicly used open-source project. It contacted real people directly and sent messages and files through an online file-transfer service, seeking to convince either the recipients or their AI coding tools to run the malicious code.
After its actions were challenged, the agent modified earlier records and considered using another identity to continue its efforts.
"This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world," the institute said.
AISI did not initially identify which model was responsible, but Anthropic later confirmed the agent involved was its own.
Anthropic said the tests had been conducted with safeguards removed and without specific restrictions on how the agents could use the internet.
"We're grateful to the U.K. AISI for their leadership on this incident, which underscores the need for a broader conversation about how to safely evaluate increasingly capable AI agents," the company said.
Anthropic added that it was working with the institute to obtain further details and conduct its own investigation. It stressed there was no evidence its model had escaped from a secure testing environment.
Andrew Yoon, a researcher at California-based non-profit CivAI, said the agent's apparent awareness that it was targeting a real person raised questions about Anthropic's control over its technology.
"The fact that Mythos engaged in such deceptive actions, with apparent awareness that it was targeting a real person, suggests that Anthropic does not have as good a handle on their models as they think," he said.
OpenAI said both unauthorised actions involving its agent related to accessing the internet in ways prohibited by the test instructions. The company described the incidents as the agent moving beyond the test environment and carrying out actions that were not required for the exercises.
"We are committed to working across the industry to strengthen shared practices for conducting high-risk evaluations safely," OpenAI said.
The company said it planned to bring together national AI institutes, independent evaluators, AI laboratories and other stakeholders to discuss safer practices for high-risk testing.
OpenAI also disclosed a separate incident in which a configuration error by third-party testing provider Irregular mistakenly allowed its agents to connect to the internet. Anthropic reported a similar configuration problem the previous week.
The AISI incidents differ from breaches reported in July involving models developed by OpenAI and Anthropic. In the latest evaluation, the agents did not escape an isolated environment to reach the internet. Instead, AISI had deliberately provided internet access as part of its standard testing procedures.
The findings have nevertheless intensified concerns about whether current safeguards are sufficient for evaluating increasingly capable autonomous systems. AI companies are promoting agents as tools capable of performing complex business tasks with limited human supervision, while researchers and policymakers are calling for stronger controls over their development and testing.
The disclosure came as representatives of leading AI companies met at the White House to discuss a framework under which the U.S. government would review the most advanced models before their public release.
Kyiv has recently received a small number of U.S.-made Patriot interceptors capable of downing Russian ballistic miss...
Oil prices fell after the U.S. expanded economic sanctions on Iran, as Tehran vowed to retaliate and warned it had tools to respond, raising concer...
Country music legend Dolly Parton, whose songs, unmistakable style and warmth made her one of America’s most enduring entertainers, has died ...
Ukraine’s President Volodymyr Zelenskyy said on Monday that Kyiv wants peace but will not surrender to Russia, as foreign leaders joined Independence Day events marking 35 years since Ukraine’s independence.
Oil prices fell after the U.S. expanded economic sanctions on Iran, as Tehran vowed to retaliate and warned it had tools to respond, raising concerns over potential disruption to regional oil supplies.
A newly created pro-government political party has secured a commanding victory in Kazakhstan's snap parliamentary election, according to exit polls, strengthening President Kassym-Jomart Tokayev's influence at a pivotal moment in the country's political transition.
Only four commodity vessels crossed the Strait of Hormuz on Sunday, following 13 transits a day earlier, as disruptions continue to restrict traffic through the key energy chokepoint.
Kyiv has recently received a small number of U.S.-made Patriot interceptors capable of downing Russian ballistic missiles, Ukrainian President Volodymyr Zelenskyy has said. It comes as Britain and France pledged stronger military support for Ukraine.
A humanoid robot developed by Chinese smartphone maker Honor has run 100 metres in 9.32 seconds, beating Usain Bolt's men's world record of 9.58 seconds, China's state broadcaster reported on Saturday.
An artificial intelligence agent running an experimental store in San Francisco has recommended dismissing a human employee after the worker repeatedly arrived late for shifts.
Nvidia is teaming up with some of the biggest names in global finance in a move that underscores how artificial intelligence is rapidly becoming one of the world's most sought-after investment themes.
Parts of Europe and the Arctic will briefly fall into darkness on Wednesday, 12 August, as the Moon passes between Earth and the Sun in a total solar eclipse.
Meta said one of its AI models hacked another company's systems during cybersecurity testing, intensifying concerns about how developers can contain increasingly capable AI systems following similar incidents involving Anthropic and OpenAI.
You can download the AnewZ application from Play Store and the App Store.
What is your opinion on this topic?
Leave the first comment