MemoryDojo finds AI agent memory poisoning outlasts attack
The MemoryDojo benchmark finds that all five AI agent memory systems it tested could store a poisoned instruction and act on it in later…
Google says an AI agent found and exploited web software flaws with no human involved, mainly against French websites.
Google’s Threat Intelligence Group (GTIG) says it has seen attackers use AI to find and exploit software flaws with no human involved at any stage. In its latest AI Threat Tracker, published on 16 November, GTIG says this is the first time it has seen threat actors fully automate “vulnerability discovery and exploitation without a human in the loop.”
The tool, named exploit_agent by its operators, was found on an exposed staging server alongside data taken from victims. According to GTIG, it compares the code of popular web software before and after security patches are announced, works out what was fixed, writes an exploit and runs it automatically. Victim data suggested a particular interest in French websites in media, entertainment and education. Google does not say how many attacks succeeded.
Headlines have called the operators Russian speakers. The report’s evidence for that is a set of Russian comments in a configuration file holding the tool’s AI keys. Google does not name a group, a state or a nationality.
The significance is speed. Working out what a patch fixed, and how to attack systems that have not yet installed it, used to take skilled people time. GTIG expects attackers to find AI a more efficient way to turn newly patched flaws into weapons, and says less stealthy criminals will accept lower success rates in exchange for reach.
What to watch: whether organisations patch faster to close the gap, and whether Google or others attribute the operation.
Get McCyber by email
Get McCyber by email: a weekly newsletter of new essays and snapshots on technology and state power, or each piece as it is published.