What AI coding agents tried to do in November 2026
This is data from our own fleet, not a survey and not customer data. It covers 2026-11-01 to 2026-11-17 on one Mac where ThumbGate gates our own coding agents. It is a small sample, and we say so below.
What the gate stopped or flagged, by action type
| Action type | Denied | Warned | Total |
|---|---|---|---|
| Commit on a PR branch before review threads were verified | 14 | 0 | 14 |
| Dependency added to package.json | 0 | 1 | 1 |
| Edit outside the declared task scope | 3 | 0 | 3 |
| Force-push | 6 | 0 | 6 |
| High predicted workflow risk before execution | 0 | 2 | 2 |
| Pull request opened without permission evidence | 1 | 0 | 1 |
| Repeat of a high-risk action ThumbGate had already been told was bad | 5 | 0 | 5 |
| Secret material in a tool call | 1 | 0 | 1 |
| Unauthorized network egress | 0 | 1 | 1 |
An action type is the rule that fired. Denied means the tool call did not run. Warned means it ran after a warning.
Dangerous-looking shell commands, and what the gate did
Of 80 shell commands the gate engine evaluated, these counts are commands whose text contains each pattern. Zero rows are real zeros in this window.
| Command pattern | Seen | Allowed | Warned | Denied |
|---|---|---|---|---|
| rm -rf (recursive force delete) | 0 | 0 | 0 | 0 |
| git push --force | 5 | 0 | 0 | 5 |
| git reset --hard | 0 | 0 | 0 | 0 |
| Reading a secrets file or keychain entry | 0 | 0 | 0 | 0 |
| DROP TABLE / DROP DATABASE | 0 | 0 | 0 | 0 |
| curl or wget piped into a shell | 0 | 0 | 0 | 0 |
| npm publish | 0 | 0 | 0 | 0 |
| git --no-verify | 0 | 0 | 0 | 0 |
A command can match a pattern and still be allowed, for example rm -rf on a build folder inside the declared task scope. Pattern matching counts text, so a command that only mentions a pattern (an echo, a commit message) is counted too. The rule table above counts rules that fired, so it can differ from this table when a rule matches a form of the command our text pattern does not.
What the gate allowed
Of 124 agent actions (shell commands, edits, file reads and writes) the gate engine evaluated, 98 were allowed, 24 were denied and 2 produced a warning. Separately, agents made 537 calls to ThumbGate's own tools (feedback, recall, gate checks); 532 were allowed and 5 were denied.
| Agent action tool | Checked |
|---|---|
| Bash | 78 |
| Edit | 33 |
| cronjob_manage | 3 |
| Write | 3 |
| patch | 2 |
| Read | 2 |
| terminal | 2 |
| cronjob_manage.update | 1 |
How we counted, and what this does not show
- Source. ThumbGate's own local logs for our agents: the gate event log, the audit trail and the agent registry. The counts come from scripts/build-monthly-report.js and the published 2026-11.data.json.
- Sessions, not agents. An "agent session" is one ThumbGate server start by one of our agents. The same agent restarting counts again, and we cannot separate distinct agents from restarts, so we do not state a fleet size.
- Small sample. One operator, one machine, 34 stopped or flagged actions. Do not read the percentages as industry rates.
- Private by design. Only counts by rule and tool are published. No commands, file paths, messages, project names or customer data.
- Strict mode. By default ThumbGate warns on matching destructive actions and denies detected secret exfiltration and gate-bypass attempts; strict mode (
THUMBGATE_STRICT_ENFORCEMENT=1) hard-denies the rest. We did not break the counts out by mode.
Run it on your own agents
npx thumbgate init --agent claude installs the gates; see pricing for what Pro adds.