Verdict reviews hostile pull requests and executes agent-supplied verification commands. That makes its trust boundary load-bearing: please report vulnerabilities privately so a fix can land before details go public.
Use GitHub private vulnerability reporting: the repository's Security tab → Advisories → Report a vulnerability. Do not open a public issue for anything you believe is a security bug.
Include the verdict version (verdict --version), the install route
(installer script, npm, Homebrew, skill copy, or source build), the
verb and flags involved, and steps to reproduce. Redact tokens and
private repository data from the report.
We aim to acknowledge reports within 3 business days, agree a fix and disclosure timeline with you, and credit you in the release notes unless you prefer to stay anonymous.
Pre-1.0, only the latest release is supported with security fixes.
Upgrade with verdict upgrade (or re-run the installer) and re-check
before reporting: many findings are already fixed on the latest tag.
In scope (highest priority first):
verify --sandbox, the enforced boundary: mount and namespace isolation underbubblewrapon Linux, network denial undersandbox-execon macOS, the credential-path hiding, and thesandbox-unavailablerefusal that never degrades to an unwrapped run. An escape from that sandbox, or a run that records"sandboxed": truewithout the enforcement, is a vulnerability.- The
verifygate's own contract:verifyexecutes the agent's claimed proof (verifyCmdviabash -c, 300s watchdog, tail recorded into the patch record). The pre-exec refusals are a tripwire, not a boundary: the gate refuses the spellings it knows, and averifyCmdspelling it does not know executes inside whatever boundary you gave it — that ceiling is documented in the manual and a hit on it is a gate gap to report and arm, not a vulnerability. What is in scope is the tripwire's own promise: a command that runs after the gate refused it, a spawn that happens before the refusal, a refusal that names the wrong reason, or a spelling the docs and the fuzz corpus claim refuses while it executes. - The installers (
install.sh,install.ps1): archive checksum, cosign verification pinned to the release's published build commit and to a push-triggered run, atomic replace, no execution before verification. A verification bypass, a path that runs an unverified download, or an install that accepts a release withoutcommit.txtis a vulnerability. - The
upgradechannel: verified self-replace for installer-managed installs, owning-manager delegation otherwise (never privilege escalation, never background checks). - The
setup-verdictaction and the skill install/bootstrap paths (install-skill.sh,verdict skill install,scripts/verdictbootstrap): same bar — no unverified execution, no writes outside installer-owned directories. - Fetched diffs and review comments are untrusted input throughout; prompt-injection handling and secret redaction in recorded tails are in scope where they cross a trust boundary.
- Control-plane integrity: the review controls (
.verdict/config.json,.verdict/kb.jsonl,.verdict/learnings.md, and.verdict/overrides/fragments) are read from the PR base commit and snapshotted into the bundle, so a pull request cannot edit the rules it is judged by.postandrecordrefuse a missing or drifted snapshot (controls-unavailable,controls-stale). Any path where head bytes steer a review without the recorded--allow-head-controlsopt-out is a vulnerability.
Out of scope: a verifyCmd spelling the string gate does not know (the
documented tripwire ceiling above — send it, it gets armed and added to
evals/fixtures/verify-fuzz-corpus.tsv), documentation typos (open a regular
issue), missing
hardening wishes with no exploit path (open a feature request), and
vulnerabilities in apps/docs build-time dependencies that never
reach the published static site (still reportable, assessed by
reachability).
Verdict spawns two kinds of child, and their environments differ on purpose. Do not "fix" the inherited side: it is the documented local posture.
- Inherited: the fixed
ghandgitargv verdict itself issues (fetch, post, triage,doctor, and the base-owned control reads). These run with the developer's environment because verdict is a local CLI acting for the operator: the calls must use the operator's existingghauth and git configuration, and shadowing either would break the command rather than harden it. The spawn site passes no environment, so the child inherits the parent's (internal/core/exec). - Default-deny allowlist: the agent-written
bash -cverify command.verifyruns the agent's claimed proof (or deep-group--command) with a default-deny filtered slice of the caller's environment (internal/core/verify/env.go), never by inheritingos.Environ()unfiltered. Only baseline locale, terminal, shell, XDG, and toolchain variables (PATH,HOME,LANG,LC_*,TERM,TMPDIR,TZ,CI,USER,LOGNAME,SHELL, color flags,XDG_*_HOME, Go/C/Rust/JS/Python cache and home variables, plus Windows OS runtime variables onGOOS=windows) pass by default; any variable whose name matchesenvsecret.IsSecret(credential-shaped endings such as*_TOKEN,*_SECRET,*_KEY,*_PASSWORD, plusAWS_ACCESS_KEY_IDandSSH_AUTH_SOCK, includingLC_TOKEN) is dropped even when it matches an allowlist prefix. An explicit opt-in keeps the named variable even when its name is secret-shaped. Operators opt additional variable names in with--env-allowlist NAME[,NAME...]or.verdict/verify-env(one variable name per line,#comments and blank lines ignored, read exclusively from the verified base-commitcontrolssnapshot, checked against the bundle's current base, so an untrusted PR head cannot widen its own environment unless the operator passes--allow-head-controls, which reads the worktree instead). Dropped variable names — never values — are reported in sorted order inwarnings[], and the full parent snapshot remains the redaction list forverifiedTailinpatch.json.
The hermetic counterpart for tests is internal/testenv: every harness that
spawns git, gh, or the verdict binary builds its child environment there.
It drops secret-shaped variables (*_TOKEN, *_KEY, *_SECRET, AWS_*),
harness knobs (REVIEW_*, GH_*, GITHUB_*, GIT_CONFIG_*), and git's own
location variables (GIT_DIR and friends, which git exports to hooks), then
pins git's configuration and identity, so a test never inherits the runner's
environment. Each git-running package's TestMain applies the same scrub to
the process, so a suite launched from a hook writes only to its fixtures.
Production inherits the operator's, tests never do — both are deliberate.
No auto-updater phones home by design, and GH_TOKEN never leaves
the operator's machine — please do not test with tokens or private data
you do not own, and do not probe live infrastructure beyond a minimal
proof of concept. Local file scope: verdict opens only files it is
explicitly handed (bundle inputs, VERDICT_USAGE, settings, stores it
owns); it never discovers or opens other tools' transcripts, logs, or
config directories.