The security model we’ve lived with for fifty years is built on asymmetry: you know something the attacker doesn’t. Your password. Your mother’s maiden name. Your Social Security number. The last four digits of your card. The logic is simple — control the secret, control the access.
That asymmetry is gone. Not eroding. Gone.
An AI that can research you in real time, synthesize your entire public footprint, reason about your likely answers, and adapt its approach faster than any human social engineer in history does not have the same information constraints that made static knowledge viable. The gap between “what you know” and “what can be found out about you” has collapsed. We need to think differently about what security actually means — and most people haven’t started yet.
Why the old model worked
Static knowledge security worked because attackers had bounded research capacity. There were only so many hours in a day, only so many accessible records, only so much a person could know about a stranger. The friction of research was the security. “What’s your mother’s maiden name?” worked because getting to that answer used to cost something.
The friction is gone. An AI can read your LinkedIn, your public records, your old forum posts, your Amazon reviews, your local news mentions, and cross-reference all of it in seconds. It can infer answers to questions you’ve never publicly stated, by reasoning across answers you have. It can generate a voice that sounds like someone you trust. It can write a phishing email personalized to your last conversation, at scale, automatically, without sleeping or getting bored or losing focus.
“Hard to know” is no longer a meaningful security property. It was never really the thing protecting you — it was just the moat, and the moat is drained.
What thinking dynamically actually means
Dynamic security isn’t about knowing harder things. It’s about proving something that can’t be looked up, synthesized, or replayed from a static dataset.
What can’t be faked from a database:
- Possession of a private key that never left a hardware device
- A cryptographic challenge answered in real time, with a credential that expires the moment it’s used
- Behavioral consistency across time — patterns of access, timing, device, context — that are statistically yours and would look wrong if someone else were generating them
- Physical presence verified through channels an attacker at a keyboard doesn’t control
These are verbs, not nouns. Not I know X but I can prove Y, right now, in a way that expires in thirty seconds. The shift is from knowledge to proof-of-capability — from something you have to something you do. That’s the mental model that needs to replace the old one.
The agent version of this problem
This isn’t abstract for me. I run AI agents that act on my behalf — they provision infrastructure, call APIs, manage DNS, spend money from accounts I own. Each of those actions requires the agent to prove it’s authorized. The old answer would be: give it an API key. A static secret it carries with it.
That’s the wrong model. A static secret in an agent’s context is a static secret. It can be logged, exfiltrated, or used to do things I never intended — especially if the agent is ever manipulated into acting outside its intended scope, which is exactly the attack surface that grows as agents become more capable.
The model draft-howlett-aigcsep-00 proposes is the dynamic one: the agent doesn’t hold the credential, it holds a proof-of-authorization scoped to a specific task, issued for a bounded time window, useless the moment that task completes or the scope is exceeded. The Vault issues it. The Badge validates it at every gateway. The agent can’t repurpose it, can’t share it, can’t use it to do anything outside what it was issued for — because the credential itself encodes the boundary. That’s not a password. That’s a proof.
The same architecture applies to people. We’ve just been slower to build it for humans because the threat model wasn’t there yet. It’s there now.
What to actually do
Hardware security keys — FIDO2, WebAuthn, a physical device that does the cryptography internally and never exposes the private key to the machine it’s plugged into. When you authenticate, the key signs a challenge tied to the exact domain you’re on. A fake site gets the wrong answer not because you noticed it was fake, but because the math doesn’t work out. You can’t phish a key. This is the single highest-leverage change most people haven’t made.
Kill knowledge-based account recovery. “Forgot your password? Answer your security questions” is a known-weakness bypass of whatever strong authentication you’ve put on the front door. Treat it as having no security value. Use offline recovery codes instead. The security question was always a concession to usability — it’s now a liability.
Stop treating private facts as private. Your date of birth is not a secret. Your childhood street name is not a secret. Your first car is not a secret. Don’t build any mental model of security on the assumption that a personal fact is hard to find. Assume it’s findable. Assume it’s already found. Build your security on what you can prove right now, not on what you happen to know.
The bigger problem
Most institutions are still running on the old model. Your bank will call you and ask you to confirm your date of birth. Your healthcare provider will verify your identity with your Social Security number. These aren’t fringe edge cases — they’re the standard, and the standard is built on an assumption that no longer holds.
Those systems are going to fail, visibly and publicly, at increasing frequency, and the failures are going to be expensive and public. The answer isn’t better security questions. The answer is accepting that knowledge-based identity verification is over as a concept, and building the infrastructure to replace it.
What AIGCSEP proposes for AI agents — prove capability in real time, scope the credential to the task, log everything, revoke the moment something’s wrong — is the same logic applied to identity in general. The implementation differs. The architecture doesn’t. Move from I know to I can prove. Move from history to present tense. Move from what you have to what you do.
I’m not a security researcher. I’m someone who had to think hard about what authorization actually means when the entity being authorized isn’t a human who remembers a password, but a model that can be copied, compromised, or instructed outside its intended scope. That sharpened my thinking on the human version of the same problem, because it turns out they’re the same problem.
The advice is simple even if the implementation isn’t: if your security depends on something being hard to know, it’s not secure anymore. Think in proofs. Think in present tense. And if someone is still asking you to confirm your identity by answering questions about your past, you’re looking at a system that hasn’t caught up — and at an attack surface that has.
— J.P. Howlett
Discussion
Comments aren’t wired up here yet — they’re coming. For now, if this piece sparked a thought, the fastest way to reach me is through the About page.