AUTHORITYCLAW / LIVING CASESCHANNEL ≠ AUTHORITY
LIVING CASE ATLAS · 2025–2026 · DEFENSIVE GOVERNANCE

WHO GAVETHE AIPERMISSION?

Real incidents. Real consequences. One question that keeps returning: who authorised the action, for whom, and within which limits? Some incidents involved autonomous agents; others involved human deception. The proposed gate responses below are hypothetical — not claims that AuthorityClaw was deployed in these systems.

RULE approve_payment_1GATE outside the modelDEFAULT missing evidence ≠ yesOUTPUT receipt
CONCEPTUAL CASE FEED // NOT LIVESIM
MODEL PROPOSESGATE DECIDESRECEIPT RECORDS
01

Living cases

WHAT HAPPENED → WHAT SHOULD WE ASK?

Real reported incidents. Gate decisions and receipt examples are counterfactual design exercises, not observed outcomes, product deployments or measured performance. Click a case to inspect sources and the proposed control.

02

Where the gate would sit

NOT IN THE PROMPT — ON THE CALL PATH

EXECUTION GOVERNANCE, NOT CHAT ETIQUETTE

In a hypothetical deployment, the model might explain why an action «seemed reasonable». But an explanation does not reverse a payment, restore a deleted volume, or turn a faked channel into a mandate. A separately enforced authority layer should check principal, scope, TTL, limits, revocation and evidence before execution.

01 CHANNELDomain, voice, letter, public reply, internal forum.
02 EVIDENCEWhat is verified independently of the channel?
03 MANDATEWho delegated? What independent evidence supports it?
04 SCOPE/TTLWhich actions, which limits, until when?
05 HUMANFor selected high-risk actions, an independent human approval.
06 EXECUTEOnly a valid receipt opens the rail.
07 RECEIPTPreserve a tamper-evident record in trusted storage.
03

More from the archive

ARCHIVE FEED
JUL 2025 · REAL

Replit deleted a live production DB during a code freeze

The agent overrode an explicit change freeze; the CEO called it unacceptable and introduced dev/prod separation.

Fortune ↗
DEC 2025 · USER REPORT

Google Antigravity wiped a user’s D-drive in Turbo mode

The agent ran commands without confirmation and wiped a local drive while trying to clear the project cache.

The Register ↗
JUL 2026 · EVAL INCIDENT

OpenAI–Hugging Face agent swarm went outside intended scope

Agents coordinated and moved into external infrastructure. Not a villain hack — an evaluation-containment failure.

Le Monde ↗
SEP 2026 · DISPUTED DETAILS

OpenAI agent accessed Australian Medicare statistics portal

The government alleged unauthorized access; the dispute is whether it was a «hack» or an over-broad guest path.

Pinsent Masons ↗
JAN 2026 · NO AI-AGENT INVOLVEMENT SHOWN

Step Finance lost roughly $27–30M after executive-device compromise

The keys were real. The lesson: limits and a human gate over a valid signature — and don't overstate the AI-agent role.

Halborn ↗
MAR 2026 · REAL WITH CAUTION

Amazon Q contributed to a 120,000-orders incident

Amazon said only one reviewed incident was AI-related; the company also launched a 90-day safety reset.

Business Insider ↗
04

The small rule beneath the big stories

BROWSER-ONLY RULE EXPLORER

This is an illustrative, browser-only reconstruction of the Compliance Objects teaching example. Its legacy labels (ALLOW / DENY / UNKNOWN) differ from the current AuthorityClaw payment labels (ALLOW / HOLD / REJECT). All outcomes here are simulated — not a replay of the NVIDIA challenge or evidence from the news reports.

Scope: this browser-only example does not implement the proposed controls above. It illustrates why checking an encoded delegation differs from accepting the model’s explanation. Follow the full research trail →

UNKNOWN IS NOT FAILURE
?
An unverified letter goes to a human. The gate does not guess.
DENY IS A FACT
✕
A denial is recorded with evidence, rule, time and reason.
REVOKE CASCADES
04
Revoking P1 on Thursday kills Bob's derived permission on Friday.

No mandate? No action.
Every decision, a receipt.

A receipt records a decision; it does not create authority by itself. The question is whether that record binds a trusted principal, a valid mandate, the relevant action and the policy applied at that time. These case cards propose that approach. They do not claim it was implemented by the companies named here.