Autonomy for AI agents is not blocked by a missing model. It is blocked by a missing guarantee: what may this agent do, and can you prove it after the fact? That is the problem we work on. We call it Proof-Carrying Authority.
@hadihosn Naming an owner is step one. Step two is making the owner's intent machine-checkable: the owner signs what the agent may do, and every action carries proof it fits that grant. Then accountability isn't rebuilt from logs after an incident.
ChatGPT Wallet asks for a fingerprint on every agent purchase. Right call for a $66 gift.
It won't scale to an agent paying 400 invoices. The scalable version: sign a bounded mandate once (payee, amount, purpose) and every payment proves it fits before money moves.
@HealthISAC Cold-chain storage is where detect-and-respond comes too late. Vaccine fridges have a known safe range. Any setpoint change, from a person, a script or an agent, should prove it sits inside that range before the controller accepts it.
@harisn The striking part: the whole loop ran on stolen keys, and nothing ever asked what a key was for. Keys bound to one task with a short expiry would have broken the cycle at the first theft.
Amazon's decoy LLM APIs caught attacker agents running ARTEX: stolen API keys paid for more scanning, which found more keys.
A key is authority with no context. If each credential named its task, scope and expiry, a stolen one would be worth nothing outside that task.
@HealthISAC For autonomous clinical robots, detection after the fact is too late. The safer pattern is a written envelope of what the device may do in each procedure, checked by a machine before every command, so a tampered instruction can't execute.
@MsftSecIntel Strongest case yet for bounds outside the model. If evals can't prove a checkpoint is clean, assume it isn't, and make every tool call show it's inside an allowed policy before it runs. A triggered backdoor then hits a wall instead of a shell.
DeepMind's agents produced machine-checked proofs for Signal's compiled code in about four weeks. Formal proof is getting cheap. That changes what we can ask of AI agents: not "is it probably safe?" but "show the proof it can't leave its policy."
Cloudflare and World's Trusted Agent Delegation asks an agent for only as much proof as the task needs, up to a unique human for high-risk actions. Smart. The next question for regulated work: proof of what this agent is allowed to do, checked before the action runs.
@RomanoRoth Your word 'predefined' is the key. If the envelope of allowed fixes is written down and machine-checkable ahead of time, an agent can act at 3 a.m. and show that its fix sat inside that envelope. Anything outside it waits for a person.
@MurrLincoln Limits the agent can't edit answer half of that question. The other half is evidence: a per-action record showing which limit was checked and that it passed. That's what an auditor or a risk committee will ask for once agentic trading reaches regulated accounts.
Coinbase just shipped server-side limits the agent can't edit. That's the right architecture.
Now carry it past trading: a claims approval, a PLC setpoint, a refund. The hard question is whether each action can prove it stayed inside the limit before it ran.
@cathy_slider Often nobody does, at least not in a form the agent can be checked against. Policy lives in documents while the agent holds a token. Until boundaries can be checked by a machine before the action, governance stays a slide.