Artificial intelligence is becoming ubiquitous, so in the coming years, LLM agents will be increasingly mentioned as “active participants” in various incident reports.
Such an example is present in the report published today by our Compromise Assessment service experts. The report is dedicated to incidents that remained undetected for weeks, months, or even years, and were only discovered in 2025 after requests to this service.
The absence of documented policies and clear data handling rules was the root cause of 24% of those missed incidents. Such a policy gap often occur when generative AI development tools are used.
During one project, our CA experts identified a macOS workstation that executed the Claude Code command-line assistant as a VS Code extension. The tool automatically captured filesystem snapshots to enrich its language model prompts. The parent command line looked like this:
/bin/zsh -c -l source /Users/[REDACTED]/.claude/shell-snapshots/snapshot-zsh-[REDACTED].sh && eval 'ls -lh "/Users/[REDACTED]/Documents/[REDACTED]/"*.xlsx' \\< /dev/null && pwd -P >| /var/folders/[REDACTED]/claude-[REDACTED]The snapshots collected by the AI assistant included full directory listings and paths to several Excel workbooks containing internal confidential data:
ls -lh /Users/[REDACTED]/Documents/[REDACTED].xlsx /Users/[REDACTED]/Documents/[REDACTED].xlsx /Users/[REDACTED]/Documents/[REDACTED].xlsx .. [REDACTED]The organization was advised to conduct awareness sessions for employees on the risk of exposing confidential data to generative AI tools, and to develop a policy governing the use of such tools.
More examples of unnoticed incidents, as well as recommendations for their detection and prevention, can be found in the full version of the report “Missed incidents, persistent threats, and response gaps”.
As for the danger of overly autonomous AI assistants on corporate devices, we have a guide on their detection and disabling: part 1, part 2, part 3.
