Attackers Are Now Using AI Agents to Run Intrusions at Machine Speed, Google Warns
Google Threat Intelligence Group (GTIG) and Mandiant report that adversary use of AI has crossed a significant threshold. Earlier research described early interest in autonomous "agentic" tools. The latest quarterly update says multiple threat actors are now abusing open source agentic penetration-testing frameworks to automate specific phases of live intrusions.
The frameworks named include CAIRN, PentAGI, CyberStrikeAI and HexStrike AI, alongside the widely publicised ARTEX. GTIG says it has seen CAIRN and PentAGI deployed in the wild, including CAIRN activity against targets such as the US aerospace sector, and has seen experimentation with CyberStrikeAI and HexStrike AI. These tools share common abilities: finding and exploiting low-complexity vulnerabilities, using AI models to generate single-use payloads on demand, and adjusting tactics based on real-time feedback.
Google stresses that the mechanics of the attacks are unchanged. Attackers still rely on compromised credentials and unpatched systems. What has changed is the speed and scale at which automation can chain these known weaknesses together. GTIG also reports that attackers are using AI models and agents to reverse engineer patches, helping them quickly weaponise known vulnerabilities in internet-facing servers without human intervention.
The report describes other trends too. Criminals practise "LLMJacking", stealing AI credentials and cloud resources from compromised environments and reselling access through underground proxy services. State-backed and commercial influence operations use AI to produce synthetic media and personas at scale, and espionage groups use AI for target research and data analysis, including manipulating coding assistant configurations to drop backdoors.
Why it matters
The patch-and-fix window many small businesses rely on is shrinking: if attackers can automatically reverse engineer a patch and exploit the flaw in internet-facing servers, a delay of days or weeks becomes a real exposure. The attack techniques are unchanged, so basic hygiene gaps such as unpatched systems and compromised credentials are what these tools exploit at scale. Organisations holding cloud and AI service credentials also face a new cost, since stolen keys can be resold through LLMJacking.
What to do
- Prioritise patching internet-facing servers as soon as fixes are released.
- Enforce multi-factor authentication and rotate any credentials suspected of compromise.
- Audit cloud and AI service accounts and API keys for unexpected usage or access.
- Review the configurations of any autonomous coding assistants for unauthorised changes.