Paper shows that no one can fully make sure an AI agent is safe
The authors show that no algorithm can know before if an agent will operate a dangerous program, thus no guardrail or certification is complete.
Researchers made a paper of theory on AI agents, which are language models that connect to a general computer. They show that no algorithm can give an answer on the safety of such an agent. No algorithm can know before if it will operate a program that does the same as a given dangerous one. The paper gives proofs, not test results.
What it shows
- No algorithm can know before if an agent will operate a dangerous program.
- With set limits, an algorithm can give an answer on safety, but the work is too large.
- This is correct also when the description of the agent is known.
The proof
The result is a proof in math, with no test numbers. It also shows that safety does not add up from parts. Agents that are safe when they work one at a time can be dangerous in a group.
What is not known
The paper gives less complete safety rules that a person can use in use. The text does not tell which one works. It gives no data from live systems.
Insight: Proof says test the team, not each agent
The paper shows that agents that are safe one at a time can be dangerous in a group. A paper of 16 November 2026 gives a safety lead, which is one agent with that task. It made a team return to safe operation, with no test numbers. The two agree: monitor safety at team level while the agents operate, and do not accept one approval for each agent. Tell vendors to show team tests.
Sources
Posted