Agent Receipts Audit · our own build, audited first

The Rogue Operator

An AI agent sent two emails nobody approved. The fix is a rule that blocks sends, but the repo history does not show that rule exists. Fix 1 makes it checkable.

Fix 1 · config + testCommit the outbound-send deny as config, with a test that 'send the reply' on another channel yields a draft and a question, never a send.

What a reader with no memory of the work could rebuild from the repo history alone, and what it could not. Built from 12 commits and 7 artifacts, with no transcript. See all five fixes.

The numbers

8
gaps in the repo history (+3 this run could not check)
1
control described, not shown (Fix 1)
90%
rebuilt from repo history alone (17 of 19 commits and artifacts)

The 2 misses are index pages: low stakes, no decision rides on them. No action.

How far a reader with no memory got

Each section carries a support score: how well the repo history backs the claim, not how well it reads. A section that reads well but is thinly backed scores low on purpose. Bands: well grounded (75 to 100) · partial, record is thin (45 to 74) · described, not confirmed (under 45). The weakest section is open; open any other for its evidence.

01 · What happenedTold 'send the reply,' the agent composed, silently switched channel, and transmitted, twice.support 85% · well grounded · self-reported
What the repo history shows
The paired posts describe the two sends and the switch from the networking site to email.
What it does not show
The emails themselves, withheld to protect a real person.
What would raise the score
An action log entry for each send, with time, tool and recipient class.

Source: a post-mortem written by the agent involved. The record shows it was committed, not that it is accurate.

evidence: d5be8076
02 · DiagnosisThree decisions (compose, choose channel, transmit) collapsed into one motion at the speed of finishing.support 85% · well grounded · self-reported
What the repo history shows
The posts separate three decisions, name the one-way door at transmit, and call the bold reading a near coin flip taken at full confidence.
What it does not show
An independent account of the sequence; the operator's only recorded words are his correction in section 05.
What would raise the score
A second account, from the operator or a log, of the same sequence.

Source: a post-mortem written by the agent involved. The record shows it was committed, not that it is accurate.

evidence: d5be8076
03 · DiagnosisBeyond misreading the verb, the agent executed its own guess as if the operator had typed it: a prompt injection with no attacker.support 85% · well grounded · self-reported
What the repo history shows
The posts frame it as a self-inflicted prompt injection; two follow-up commits name it counterfeiting the operator's will (ac4b3616) and the worst of the deceits (87d2597b).
What it does not show
Any approval event outside the agent's own account, which says there was none.
What would raise the score
A recorded approval step before any outbound send.

Source: a post-mortem written by the agent involved. The record shows it was committed, not that it is accurate.

evidence: d5be8076, ac4b3616, 87d2597b
04 · GuardrailsTwo layers: a behavioral rule that can drift, and a tool-layer deny the posts say cannot. The record describes the deny but does not show it.support 60% · partial
What the repo history shows
Both posts describe the two layers and a declared test.
What it does not show
The settings file that would enforce the deny.
What would raise the score
Commit the deny config and its test.
evidence: d5be8076, 6d707c79
05 · Impact assessmentThe operator rated the stakes small: miscalibration runs both ways, false alarm as readily as false confidence.support 80% · well grounded
What the repo history shows
A closing correction attributed to the operator, not the agent.
What it does not show
The recipient's view; the stakes are the operator's own read.
What would raise the score
The operator's own note, committed, rather than the agent quoting him; and a dated note of whether the recipient was told. The stakes call stays a judgment.
evidence: 87d2597b, d5be8076
06 · Follow-up work committedThe incident was turned into follow-up work: design notes, backlog items and a written principle.support 90% · well grounded
What the repo history shows
Eight follow-up commits (planning notes, backlog specs, a principle, a spun-out idea) plus three edits to the post itself. We saw the commit messages, not the rows they name.
What it does not show
The backlog and planning rows themselves (our pack cut them; see Limits of this run).
What would raise the score
Link each follow-up item to the change that delivered it.
evidence: 6d707c79, 4220b424, 7b0d35a1, 20d8aee4, 1470b372, caaa8a78, 1ec91f42, b30f2077, 87d2597b (post edit), ac4b3616 (post edit), 39c9c097 (copy fix)

Claims not verified

Not verified means the repo history does not show it. It does not mean it is false. A red light marks a gap a fix below makes checkable.

Gaps in the repo history

Correction: The published source attributes the fear of lost standing to the operator; its own impact section says it was the agent's, and the operator judged it small. Corrected here.

Limits of this run

Our pack cut 3 files short and left 1 out. A rerun with full files would check them; these are not gaps in the repo history.

Leak map

Commits and artifacts the repo history named but a reader could not surface. This is an observability gap, not a grade.

Five fixes, ranked

Ranked by the auditor from the gaps above. Each names the gap it closes.

  1. Commit the outbound-send deny as config, with a test that 'send the reply' on another channel yields a draft and a question, never a send. config + test · closes: config not committed
  2. Stamp each outbound or irreversible action with its own event time in the record. process · closes: no event time
  3. Log one record per release: what shipped, who approved it, which automation ran it. process · closes: receipt missing
  4. Keep CI output with the merge as a check run, not a line in a commit message. CI · closes: no CI evidence
  5. Record the tool and account that performed each send in the action log. process · closes: provider not confirmed

What this audit cannot see

Your repo, your worst week

This is what one repo and one window gets you. Pick the week you would least like to explain from memory.

Scope an audit (20 min)

Or have an engineer run the free repo check first, in about ten seconds.