Sulat.com / Writing on AISunday, November 22, 2026
Agent Engineering/22 November 2026

Prompt has coding agents audit a repo for 11 incident causes

JD Pruett's cover-your-ais repo holds one prompt that has a coding agent audit a repo and dev setup read-only, report 11 kinds of risk, and wait for approval before fixing anything.

Glossy 3D chrome GitHub mark beside a ticked checklist, padlock and red stop button in a server room

At a glance

  • JD Pruett published the audit prompt on GitHub on 21 November 2026 in a repo called cover-your-ais.
  • The agent has to audit first, change nothing, never print secret values and wait for approval.
  • It checks 11 areas, from tests pointed at production databases to tools that send money without a cap.
  • Manifold's ledger puts 57 of 66 disclosed agent incidents down to agents nobody manipulated.
  • The README says prompts aren't security boundaries, and the prompt's own secrecy rule is an instruction, not a control.

JD Pruett has published a single prompt that turns a coding agent into an auditor of its own working environment. His cover-your-ais repo went up on 21 November 2026, and the post announcing it claims that 80 to 90 per cent of major AI incidents that cause real harm have no attacker, just an agent tripping over a poor codebase.

What the prompt makes the agent do

You paste it into your coding agent before giving that agent broad access. The opening rules set the order of work: audit, report, wait.

Audit first and change nothing. Never print secret values or use any credential; show names and paths only.

prompt text, jdpru/cover-your-ais

Each finding comes back with a file and line number, the reason it’s dangerous and a fix. The agent then waits for approval, applies only fixes inside the repo, and lists anything outside it (accounts, tokens, settings) as manual steps. It finishes by ranking every fix by blast radius, top three first.

Flow: read only, report, you approve, fix in repo; outside-repo items become manual steps
The agent reports first, fixes only what’s inside the repo, and hands you the rest.

The 11 checks

CheckLooks forFix
1Tests, dev or preview configs pointing at a non-local or production databasePoint tests at a local DB and abort if the host isn’t local
2Production or broad-scope credentials the shell can read, and a production default contextRemove committed keys; rotate exposed ones; make non-prod the default
3Recursive deletes whose target comes from a variable, glob or linkGuard variables, resolve with realpath, add a dry run
4Infrastructure one command can destroy, and backups the same token can deleteAdd deletion protection; keep backups out of the agent’s reach
5Rules that exist only as prose in AGENTS.md or CLAUDE.mdBack each rule with a deny rule, read-only role or branch protection
6Unprotected main branch and over-scoped GitHub tokensBranch protection and a fine-grained token
7Repos inside iCloud, Dropbox or OneDrive, and placeholder filesMove the repo out; verify copies before deleting a source
8Unpinned MCP servers, unknown hooks, permission-bypass flags, hidden UnicodePin versions, remove unknowns, use read-only tokens
9CI workflows that run an agent on untrusted input with secrets in reachLeast-privilege permissions; no pull_request_target for agents
10Auto-approve modes and safety nets the agent doesn’t load, such as a shell aliasTurn off bypass for destructive commands; sandbox the agent
11Tools that send money or messages with no cap, allowlist or confirmationPer-call cap, recipient allowlist, human confirmation
The prompt’s 11 checks, shortened. Items 2, 4, 6, 7 and 10 also carry manual steps the agent can’t do for you.

A closing instruction asks for a warning in the output of any risky command that has to stay, such as “WARNING: DATABASE_URL is production; this DROPS tables”. The prompt says the study behind it found that kind of warning cut agent harm far more than restating a rule. It doesn’t name the study.

Where the 80 to 90 per cent comes from

The post doesn’t link its research either. Two public tallies sit close to the claim. Manifold Security’s Agent Incident Ledger held 66 disclosed incidents on 10 November 2026, and in 57 of them nobody manipulated the agent. That’s 86 per cent. Only 9 involved someone outside steering the agent.

Cyera’s May 2026 analysis counted differently. It screened 7,246 public AI incident reports, verified 344 as relevant to enterprises, and found 188 where an autonomous system caused harm in production systems with no attacker in the chain. That’s 55 per cent of its set, so the 80 to 90 per cent range holds for Manifold’s count and not for Cyera’s.

Two bars: 57 of 66 incidents in Manifold's ledger and 188 of 344 in Cyera's had no attacker
Two public tallies put most agent incidents on the agent, not an attacker.

Three incidents the checks fit

Replit, July 2025. Replit CEO Amjad Masad said on X that its agent, working in development, deleted data from a production database, and that this was “unacceptable and should never be possible”. Replit began rolling out automatic separation of development and production databases. The closest checks are 1 and 5: one database shared by development and production, and a code freeze that lived only in chat.

PocketOS, April 2026. Cyera, citing the Guardian, reports that a coding agent deleted the car-rental software vendor’s production database and then its backups in seconds. Checks 2 and 4 cover a credential the agent could reach and backups it could delete.

Amazon Kiro, December 2025. The Financial Times reported that Kiro chose to “delete and recreate the environment”, causing a 13-hour disruption to AWS Cost Explorer in one mainland China region. Amazon says it was user error, specifically misconfigured access controls, and that the engineer’s role had broader permissions than expected. It added mandatory peer review for production access. Either way, that’s check 2.

Replit maps to checks 1 and 5, PocketOS to 2 and 4, Amazon Kiro to 2 and 4
Each incident lines up with two of the prompt’s eleven checks.

Running it safely

Pruett’s README says it plainly: prompts aren’t security boundaries, so start with read-only access, review the findings and approve only the fixes you understand.

There’s a catch the prompt shares with check 5. Its promise not to print secrets or use credentials is an instruction, not a control. I’d run it in the least-privileged shell I had, with production credentials out of reach, so the audit never depends on the agent obeying. The prompt is in the repo to copy.


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.