TL;DR - AI agents now act inside development pipelines, and many teams lose track of what those agents did. Palo Alto Networks' 2026 State of Cloud Security research found that 40% of organizations are unable to fully reconstruct an AI agent's actions, and 53% record agent activity under a human developer's name. Give every agent its own identity, short-lived keys and a log of every action.
30-second version: "When an agent borrows a developer's identity, its actions look like that developer's actions. When something breaks or leaks, the trail goes cold. 1 identity per agent, keys that expire fast, and a log of every action fix that."
What happened
Palo Alto Networks published its 2026 State of Cloud Security research on AI agents and cloud risk. The findings:
- -40% of organizations are unable to fully reconstruct an AI agent's actions after execution.
- -53% log agent activity under human developer identities or personal access tokens (a personal token is a password-like key tied to 1 person).
- -99% rely on agents for autonomous or semiautonomous actions in their development pipelines.
Why it matters
An audit trail answers 3 questions: who acted, what changed, and when. When an agent runs under a person's token, the answer to "who" is wrong, and the answer to "what" is mixed with that person's own work.
- -Incidents take longer. A deleted table or a leaked key has no clear owner.
- -Access grows quietly. An agent with a developer's token inherits everything that developer can reach.
- -Reviews lose evidence. Security and compliance reviews depend on a log that matches reality.
How it works: 3 habits
#1 Give every agent its own identity. Create a separate service account (a non-human login) for each agent, named after the agent and its owner.
#2 Use short-lived keys. Issue agent credentials that expire in hours, scoped to the 1 repository or service the agent needs.
#3 Log every action the agent takes. Record the agent identity, the tool it called, the target and the time. Keep metadata by default, so the log itself holds zero secrets.
What's next
- - This week: the team lists every AI agent and its owner.
- - This month: each agent gets its own identity and short-lived keys.
- - Tomorrow: the next short in this series covers a flaw hiding in repository settings files.
"Name the agent, log the deed; the audit trail is all teams need."
Sources
- -Palo Alto Networks, 2026 State of Cloud Security report (AI agents and cloud risk): https://www.paloaltonetworks.com/blog/cloud-security/2026-state-of-cloud-security-report-ai-agents-cloud-risk/
Disclosure: I build Pretense, a local proxy that swaps API keys, tokens and code identifiers for stand-ins before an agent's request leaves the laptop. Identifiers come back byte-exact; secrets stay as stand-ins in the reply. It keeps agent prompts and logs free of live secrets, and it works alongside habits 1 to 3.