Co-founder and CEO, Zyras (@zyrasproof). Building the authority layer for autonomous AI. Previously VP at Forcepoint, Lookout, and Symantec. Views my own.
@RomanoRoth Revocation as the last gate is right, and the survey shows why it can't be the only one. Scope each grant to one task with a short expiry and there's less to revoke: the authority lapses when the task ends. A slow revoke then stops meaning a day of unearned autonomy.
This is the question I keep hearing from CIOs in regulated industries. Approval per action is safe and doesn't scale. No approval scales and isn't safe. A mandate the agent must prove it's inside is how we get both. https://t.co/86JMu1cAmQ
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.
California says OpenAI lost control of test agents at least 80 times, including a swarm that spent days inside Hugging Face.
The lesson for every enterprise: a sandbox is a wall, not a rule. Agents need limits they carry everywhere, checked before each action.
@16vchq Free-range agents roam inside a fence you can prove. We formally verify what an AI agent can never do, so enterprises can take the human out of the loop with confidence.
@Math_files Self referential paradox. The right answer (33%) is not among the available choices, so the likelihood of picking that is 0% which collapses the probablity space to ⅓ = 33%.
@nyike@star_cio The way out of Dr. No is a conditional yes that can be checked: an agent ships when you can state what it may touch, what it can't, and who owns it, and a system enforces that before each action. Sprawl is manageable when every agent carries its limits.
@NatlCIOReview More agents shouldn't mean more approval prompts. If each agent's allowed actions are written down and checked by a machine before it acts, people only review what falls outside that envelope. Approval fatigue usually means the boundary was never defined.
Attackers run autonomous agents now too. Faster detection alone won't be the defender's edge. The edge is making every credential worthless outside the one task it was issued for. https://t.co/vZRMyGJArW
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.
Fed Governor Barr says big banks already see AI agents trying to act in customer accounts without verified consent.
That's the real question for regulated industries. Not 'is this an agent?' but 'who delegated it, for what, and can it prove that before money moves?' That's what we're building at Zyras: authority an agent carries and a bank can check before the action runs.
@SawyerMerritt now this is how you shutdown an armchair analyst doling out free "insights" on social media. What SpaceX has achieved and continuing to aim for is nothing short of remarkable.
@gdb Finding and patching bugs faster is half of it. As red-team agents get access to real networks, the other half is proving each one stays in scope: which hosts, which actions, for how long. That's what lets a CISO say yes to running them.
@liran_tal Tell me, before an agent runs, what it can actually touch: which repos, secrets and commands, and what it provably can't. Scanners tell me what went wrong afterward. I'd want the boundary written down and checked first.
IBM just bought AuthMind to watch how agents use credentials. Microsoft's red team showed backdoors can hide in a model and fire tool calls. Same lesson: controls belong outside the model. Watching is half of it. The other half is deciding what an agent may do before it acts.
@QwQiao There is no replacement for founder-level passion and the will to change outcomes. Jobs, Jensen , Musk - these are some of the greatest this generation has seen. I hope the board and the industry is kind to Jensen.
@liran_tal Maybe the agent shouldn't hold the password at all. Let a broker hold it, and have the agent ask to do one specific thing, such as 'log in and download this statement', inside an allowed window. The secret stays put, and what the agent can do stays narrow.
@davideoks The credentials still worked, and that's the real lesson. A credential says who you are, not what you're allowed to do right now. If authority were tied to a task and expired with it, the leftovers would have been useless.
Patients are already using personal AI agents to try to book appointments, Emory's chief AI officer says. Health systems now have to choose: block those agents, let them in, or build their own automated scheduling tools instead. Emory saw traffic rise this summer but hasn't measured its volume yet. https://t.co/LUDRwpCpJd
United says agents can't book flights. Emory's Chief AI Officer says patients' agents are already trying to book appointments. Everyone is framing it as block or allow.
There's a third option: let an agent in when it can show who delegated it, what it may do, and what it can't do. Being able to reach a website doesn't mean you have permission to act on it.