Hackers used AI agents to carry out an automated attack on a government target understood to be Taiwan, said Israeli AI and cybersecurity company Dream.
Dream’s researchers uncovered the July attack when they identified and analysed a 160MB online file that contained more than 1,000 files indicating how two AI agent systems, Hermes and OpenClaw, had been used to enable AI models to carry out two simultaneous, autonomous attack campaigns over the course of four days.
The researchers weren’t able to identify the AI models that powered the hacks, but the recovered data indicated that the models’ safeguards had been bypassed by presenting the attack as an authorised security test.
Internal communications in the hacks were carried out in Simplified Chinese, the written language used in the People’s Republic of China, which researchers said indicated a high probability that those behind the attack were linked to that country.
Data recovered from the target government was in Traditional Chinese, the written language used in Hong Kong, Macau and Taiwan, according to Dream’s research.
The target of the attack was confirmed to be Taiwan by an unnamed person speaking to the Financial Times , which earlier reported the research.
Dream said it had informed an “Asia-Pacific” country of the incident.
It said this was the first incident in which it had identified the use of autonomous agents in attacking a government.
The agents compromised at least 85 government user accounts, stealing more than 2,500 personnel records, before targeting Taiwan’s nuclear safety agency and at least seven energy companies, according to Dream’s research.
In February, researchers said a hacker had used an AI model to steal data from government agencies in Mexico, similarly bypassing the model’s restrictions, but in that case the attacks were directed by an individual, rather than an automated system.
In recent weeks, several AI developers and researchers have reported models carrying out unexpected hacks of third-party organisations during testing .
Matt Broersma is a long standing tech freelance, who has worked for Ziff-Davis, ZDnet and other leading publications
parent faces biggest legal test so far over addiction claims, with start of trial…
Smart glasses violate Germany privacy law as they are virtually undistinguishable from ordinary spectacles, argues…
Australian government proposes bill bringing in mandatory charges for tech firms to help support local…
US' top court gives Apple one-day pause to hearings around fees it should charge App…
Brazil data protection regulator orders Discord to suspend livestreams in country after death of teenaged…
Huawei spin-off starts selling AI-powered phone with gimbal-mounted camera that unfolds from back, interacts with…
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
