AI in the Hands of Cyber Threat Actors

Autonomous sequences are documented, and incidents in evaluation environments have shown the approach is technically feasible against real organizations. No threat actor, however, has been observed deploying a complete autonomous pipeline against real targets in vendor telemetry. That telemetry does not cover self hosted models.

Not on the available evidence. The widely cited 89% increase in attacks by AI enabled adversaries measures actors who use the tools, not a causal effect. The inflection the corpus does document is the compression of time.

Within days, and before the patch. Exploitation precedes the patch by seven days on average, and 88% of exploitations of vulnerabilities with a public proof of concept happen within 48 hours.

Like a privileged one. An agent with write permissions needs its own identity, a named owner, least privilege, logging and an immediate stop. In a Wavestone panel of around thirty large organizations, around 15% apply IAM to their AI systems, and fewer than one in ten forward the logs of critical AI uses to the SOC.

That is not established. Their vulnerability discovery and exploitation capabilities have been demonstrated in evaluation, but no offensive use by a threat actor is documented. The effect observed among attackers comes through scaffolding and publicly accessible models.

Not across the board. Citalid adjusts a single parameter at this stage, the susceptibility of exposed assets. The other effects are treated as trajectories to stress test rather than certainties to price.