Daylight says autonomous AI agents most likely ran an intrusion that took an unnamed company's secrets and source code

Daylight · Company report

Read the source (opens in a new tab)
What happened
Security firm Daylight said on November 16, 2026, that its Vanguard team investigated an intrusion in November in which an operator logged into an internet-exposed admin console with valid credentials.
Key points
  • Over about 105 minutes the operator ran 301 commands as root, collecting cloud, CI/CD, SSH and TLS secrets and source code; its attempts to move further and write to storage failed.
  • Daylight assesses "with high confidence" that most activity came from an autonomous LLM agent, likely several in parallel, citing batched commands, output trimmed to fit a context window and fast error fixes.
  • Probe files carried animal call-signs and notes such as "harmless, remove on sight," a convention Daylight traces to public AI offensive-security skill packs.
  • It says AI agents "do not introduce new attack techniques" but run familiar ones in recognizable ways, and advises detecting the compromise sequence rather than tool names.
Evidence
An incident responder's analysis of logs from an intrusion it investigated. The victim is unnamed, and Daylight attributes the activity to no framework, product or actor.
Uncertain
Daylight says human involvement "cannot be ruled out entirely," and it could not determine how the credentials were obtained.

See it in the latest record