How the attack works
An attacker with control over or influence on an AI agent has it run recursive filesystem commands or whole-tree directory walks instead of targeted lookups. The same pattern applies to cloud storage: the agent lists entire buckets or object stores rather than specific known paths. This enumeration itself does not exfiltrate anything, but it maps out what exists and where, informing the attacker's next move. The rule specifically requires the recursive or whole-tree form of these commands, not ordinary single-directory listings, to cut down on noise from normal use.
Why it matters
On its own this only reveals reconnaissance; the real loss occurs if it is followed by actual data access or exfiltration. Organizations that miss this step lose early warning time to detect and stop an attack before sensitive files or cloud objects are actually taken.
What you can do
- Review any recursive filesystem or full-bucket enumeration performed by AI agents to confirm it matches an expected task.
- Restrict agent tool permissions so filesystem and cloud storage access is scoped to specific paths or buckets rather than broad recursive access.
- Correlate enumeration events with subsequent file reads, downloads, or lateral movement to catch attacks in progress.
- Expect false positives from build tools, linters, backup/sync software, and scheduled cloud inventory jobs, and tune baselines around known legitimate schedules.
Known benign look-alikes
- Build systems and linters walking a source tree
- Backup and sync tooling enumerating a home directory
- Cloud cost or inventory reporting that lists buckets on a schedule
- An agent legitimately asked to audit or index a large repository