A daily puzzle where you trick the AI into spilling a secret. 🔓
One new puzzle every day. Talk your way past the locks and get them cooked.
Play by yourself or with a friend.
Play free: https://t.co/xTleL7t8wk
Watch: https://t.co/0Y58cV8tvx
5/
Small model of the idea, not Workday or https://t.co/HcolKGe0w9 internals. Agents, systems, scopes and times are example inputs. The deal hasn't closed yet.
Code + write-up: https://t.co/k8VgLIGARX
3/
Demo: close-helper reads finance at 09:40.
At 09:45 it moves to a support role. Finance is revoked on the spot. CRM read stays, because the new role still allows it.
A time-boxed CRM grant from 09:00 expires at 11:00. Denied after that.
4/
Then the agent ends.
Ticket world: the record says ended, but the email app still allows send.
Lifecycle world: the same app says no. The revoke reached the grant.
A retired agent with live credentials is just a ghost with an API key.
2/
The lifecycle is the design:
1. Register the agent with an owner and a role
2. Grant per system, with scopes and a reason
3. Reject anything above the role's ceiling
4. Check every action: record, status, grant, scope
5. Role change trims. End revokes everything
1/
Workday announced Oct 13 it plans to acquire https://t.co/HcolKGe0w9, an agentic identity platform for agents.
The idea: an agent gets access when it's registered and loses it the moment its work or role ends.
I built a tiny access grant for that. TypeScript, no API key.
Same 300 example contracts. Same answer (43 auto-renew).
No budget: $4.729 at peak 300 concurrent
Capped: $0.369 at peak 16
Tight $0.30 cap: refuses the merge, picks plan B, finishes at $0.236
Hard refusal when it can never fit.
Soft wait when in-flight workers will free the room.
The interesting line is the refusal:
needs $0.139, only $0.070 left.
Anthropic's docs list 64 threads at once and 1,000 agents per run.
Model requests count toward the session budget.
A run has no price of its own.
So who decided it should fan out that wide?
Anthropic's dynamic workflows let an agent hire hundreds of subagents in phases.
The model is no longer only doing the work.
It is deciding how many workers to hire.
That needs a budget, not hope.
Google released Gemini 4 Argon on Oct 13. Past the per-request limit, the API returns finish_reason CONTINUATION and a continuation token.
You resend the same request with the token, and it keeps going.
I built a tiny continuation loop for that. No API key.
Demo 3: the stream drops mid-segment.
The naive loop ends with 107 tokens for an 87-token answer. The real loop discards the uncommitted 20, retries from the same token, and matches exactly.