ESET says UAC-0099 used a GuardBreaker prompt in an attack on a victim in Ukraine, marking day 3 of agent-as-attack-vector coverage and showing that AI safety filters are now being targeted on purpose. The group hid the trigger text inside a malicious VBScript comment, where it was meant to trip the model’s guardrails and stop further inspection.
The trick is simple: a scary-looking string makes the large language model refuse or narrow its analysis, so the rest of the file can go unseen even if it is malicious. That turns the model’s own safety layer into an evasion layer and creates a blind spot after the trigger text, not before it.
For teams using LLMs to triage code, scripts, or attachments, the exposure sits in the inspection workflow itself. If the model stops at the wrong line, the file may still need a human or a different tool to see what the guardrail prevented it from reading.
The state of AI for security: Measuring what matters most for building trust | Amazon Web Services
Security teams are starting to actively use AI for security work, including vulnerability triage, penetration testing, threat modeling, incident response, and code review. The promise is speed, but a security tool that moves fast and raises too many false alarms doesn’t save time. Engineers spend time on false alarms, on-call is noisier, and teams distrust […
Introducing the CyberAgents Exchange AI Inspector: Rigorous review for community-built AI
The CyberAgents Exchange AI Inspector adds a rigorous security review, giving teams a clear signal on which AI agents, skills, MCP servers, and multi-agent playbooks are safe to bring into their environment.