Repository navigation
Releases
v0.11.1
A patch release for one false positive that a sweep of Inspect AI turned up. ACR-APP-003 (dynamic code execution)
reported 1,124 findings there, nearly all of them calls to Inspect's own eval() function.
It now skips a Python eval or exec that the file imports, defines or assigns itself (such as
from inspect_ai import eval), and a call shown inside a string, a docstring or an f-string. A call to the builtin
still fires, and JavaScript is unchanged. On Inspect AI the rule went from 1,124 findings to 8: seven call the builtin
and one is a JavaScript template string.
v0.11.0
A new supply-chain rule and a false-positive pass over seven more Go and Kubernetes projects (KubeRay, Kueue, Kubeflow
Trainer, Volcano, KAI Scheduler, llm-d, production-stack): 1,984 findings before, 1,942 after: 35 CRD schema
fields read as pod volumes, five on minified or vendored code, and two the authors had already reviewed with a
//nolint:gosec comment. Nothing new appeared. What the sweep left is mostly deployment hygiene those projects have
not adopted (unpinned actions, secrets: inherit, containers that run as root), which is what the rules are for.
Added
- ACR-DEP-006 (low): a Dependabot update whose shortest wait after a release is under seven days, whether it has no
cooldown, a shortdefault-days, or a shortsemver-*-days. Withoutdefault-daysDependabot waits three days,
and the default stops mattering once all three semver levels are set; both are read. A release that is compromised
is usually found and withdrawn within days. Security updates ignore the cooldown..github/dependabot.ymlis a
file kind of its own (dependabot). The rule flags every update in Kueue (5), KAI Scheduler (6) and Trainer (17). acr review --allreviews a directory that is not a git repository, such as an unpacked release: it walks the tree,
leaves out.git,node_modules,__pycache__and virtual environments, and never follows a symbolic link. The
other modes now sayis not a git repository; use --allinstead of showing a git command line.
Changed
//nolint:gosec,// #nosecand# noseccount as a reviewed decision for the rules those tools also check (eleven
and fourteen application rules); a bare//nolintor another linter does not.third_party/is left out by default besidevendor/andnode_modules/, and a file with a line over 20,000
characters (a minified bundle) is reported as skipped. KubeRay's swagger-ui bundles produced four findings.- Kubernetes rules skip CustomResourceDefinition documents, decided per document so a bundle that holds CRDs next to
workloads keeps its findings on the workloads. A CRD's schema nameshostPathwithout being a pod: 35 findings.
v0.10.0
[0.10.0] - 2026-11-02
Go parity, second round: the weak-hash, deprecated-TLS, JWT, CORS and cookie rules now read Go, each swept over nine Go
repositories before shipping and each leaving alone what those repositories showed to be legitimate (a tls.Config
without MinVersion on Go 1.22+, a token parsed to inspect its own claims).
Fixed
- ACR-DOCKER-012 no longer reports
pip install ".[extra]": a local path with an extra, quoted or not, is the image's
own code, pinned by the build context.
Changed
- ACR-APP-007 (weak hash) and ACR-APP-014 (deprecated TLS version) cover Go:
md5/sha1NewandSum, and a
MinVersionof SSL 3.0, TLS 1.0 or 1.1. Atls.ConfigwithoutMinVersionis not reported: Go has defaulted to
TLS 1.2 since 1.22. Go tests are skipped; they hash fixtures and pin old versions to test their rejection. Swept
over Argo CD, KServe, KAI, etcd-operator, llm-d and Tesla's two repositories: four and one findings. - ACR-APP-008 (JWT accepted unverified) and ACR-APP-009 (CORS wildcard with credentials) cover Go: golang-jwt's
UnsafeAllowNoneSignatureTypeandSigningMethodNoneoutside tests; rs/corsAllowedOrigins: {"*"}or
gin-contribAllowAllOrigins: truebesideAllowCredentials: true, and the two headers written by hand.
ParseUnverifiedis deliberately not reported: a sweep found Argo CD using it seventeen times to read a verified
token's own claims, which is inspection, not acceptance. - ACR-APP-015 (session cookie without Secure or HttpOnly) covers Go: an
http.Cookieliteral named like a session,
auth, token, sid or jwt cookie that does not set both flags to true (Go's zero values are false). Tests are skipped.
v0.9.0
[0.9.0] - 2026-11-02
Measured with scripts/sweep.py over Argo CD, Kubeflow Pipelines, KServe, MLflow and flytekit, the built-in rules'
findings went from 4,654 to 2,315 since 0.8.1; every removed finding was a test input, a fixture, the name of a secret,
or a pull request's own base branch. What a rule still reports in those repositories is what it is for.
Added
scripts/sweep.pyruns the built-in rules over whole repositories and prints findings per rule and repository, a
sample of the lines behind each rule, and with--beforehow the counts moved since the last sweep. CONTRIBUTING
asks for a sweep before a rule ships.
Changed
- ACR-K8S-005, 007, 009 and 021 (image tags, limits, non-root, restricted-profile gaps) skip manifests under
testdata/andfixtures/: inputs a project's tests read, never applied to a cluster. Argo CD alone carried
600 such findings. Manifests underexamples/,test/e2e/and the like are still read. ACR-K8S-010 and 020 (RBAC
wildcards and escalation paths) skip the same directories: a ClusterRole fixture is never bound. - ACR-AI-004, 005 and 010 (unpinned model loads, LLM calls without a token limit, agents run unsandboxed) skip test
modules, which load small public models, call mocked clients and assert that the dangerous flag is absent.
MLflow's findings for the three went from 186 to 61. - ACR-GHA-002 (untrusted event data in a run step) no longer reports
github.event.pull_request.base.refor
base.sha: the base branch is one of the repository's own, chosen from a list. The head ref is still reported.
v0.8.1
[0.8.1] - 2026-10-31
A false-positive release. The built-in rules were swept over fifteen repositories (Argo CD, Kubeflow Pipelines,
KServe, MLflow, flytekit, KAI Scheduler, etcd-operator, Tesla's vehicle-command and fleet-telemetry, llm-d's router
and simulator, and this project's siblings); everything below came out of that sweep. The hardcoded-secret rule's
findings on those repositories went from 1,556 to 101.
Fixed
- The pod rules no longer crash on a manifest whose
containers:is written as a mapping, or stands in as a template
placeholder; the review reported an internal error for such a file (seen on a kruise DaemonSet fixture in Argo CD). - ACR-K8S-013 (hostPort) no longer reads kind cluster configs, whose
extraPortMappingscarryhostPorttoo: a YAML
file whose documents all name a kind and none is a workload is skipped. Templates and fragments are still read. - ACR-APP-001 (SQL built with string formatting) skips test files in every language: tests build statements from
literals to set up and inspect rows, and the statements that matter are in the code under test. - ACR-DOCKER-012 says a
pkg>=1.0install has a range, not that it has no version. - ACR-SEC-002 (hardcoded password or token) no longer reports where a secret lives or comes from: names ending in
name,ref,key,endpointand the like, values that are lower-case words joined by separators
("argocd-initial-admin-secret"), the UPPER_SNAKE name of the environment variable a secret is read from, and
URLs, camelCase identifiers, a shell variable reference and a value that is the secret word itself
(password = "password"names a field). Across Kubeflow Pipelines, MLflow, Argo CD, KServe and flytekit its findings went from 1,556 to 101; what
remains mixes letters with digits or symbols, as credentials do.
v0.8.0
v0.8.0
Go is now reviewed as fully as Python: the SQL, shell, TLS, open-redirect and path-traversal rules cover Go's
standard library and its common web and database packages, and three Go-specific rules join them. Every rule here
was run over llm-d, Argo CD, KServe, Kubeflow Pipelines, MLflow and etcd-operator before shipping; the false positives
that turned up (an HTTP client's Get read as SQL, TLS turned off in tests, a umask-masked mode) are fixed.
Added
ACR-APP-024: Go servers started withhttp.ListenAndServeor anhttp.Serverwith noReadHeaderTimeout, open
to Slowloris. Argo CD has eleven, Tesla's vehicle-command HTTP proxy two.ACR-APP-025: files and directories created world-writable (0777, 0666) in Python and Go.ACR-APP-026: an authorisation or identity check written as a Pythonassert, which-Oremoves.ACR-GHA-018(high): an action whose owner is one edit, or one swapped pair of letters, from a well-known one
(action/checkout,actoins/setup-python,dockr/login-action).ACR-K8S-021(low): containers withoutallowPrivilegeEscalation: falseorcapabilities.drop: [ALL], the two
restricted-profile settings every container needs.ACR-DOCKER-012:pip,npm -gandgo installof packages with no version in a Dockerfile, one finding per
RUN. Argo CD's and Kubeflow's build images install several tools at@latest.ACR-APP-022:tarfileandzipfileextraction with nofilter=and no member-path check, the zip-slip shape of
model and dataset downloads. Run over MLflow, KServe and Kubeflow Pipelines before shipping: four real
extractions, no false positives left.ACR-APP-023: temporary files created by a fixed name under/tmp(tempfile.mktemp, a literal joined onto the
temp directory,open("/tmp/x", "w"), Go'sfilepath.Join(os.TempDir(), "x")), open to a symlink race. Over Argo CD it found the dex
command writing its TLS key and config to fixed paths under/tmp, and MLflow's REPL cache under/tmp/repl_tmp_data.ACR-DEP-005: pip requirements installed from a git branch, tag or bare URL instead of a commit.
Changed
ACR-APP-010(predictable random numbers for a secret) covers Go'smath/rand, in files that do not import
crypto/rand.ACR-APP-001,-002,-005,-012and-013cover Go: statements built withfmt.Sprintfor+for
database/sql, sqlx, pgx and GORM;exec.Command("sh", "-c", ...)with a built command;InsecureSkipVerify: true
andgrpc.WithInsecure();http.Redirectand Gin redirects to request values; file calls and path joins fed by
request values, route variables or form fields.ACR-APP-005skips test files, where verification is turned off against the test's own self-signed server.
Fixed
ACR-K8S-005no longer reportsimage: ${VAR}or$(VAR)placeholders as unpinned. The image is filled in at
deploy time, often from a file of digests, so the pin is not in the manifest; the Compose rule already skipped them.
v0.7.0
v0.7.0
Lock files are now reviewed, and so are HTTP calls with no timeout, the checkout token shipped in a workflow
artifact, and RBAC escalation paths short of *; 118 rules in all. The lock-file and Go key rules come from real
failures this month: a CA bundle that kept trusting a compromised root, and a timing side channel in a Go key check
that a maintainer caught in review. The timeout rules came out of reading LLM-serving clients' HTTP code, the artifact
rule out of Unit 42's token-leak research, and the RBAC rule out of the Kubernetes RBAC good-practices page. Each new
rule was run over real repositories before it shipped, and the counts are below.
Added
- A
lockfilefile kind:uv.lock,poetry.lock,Pipfile.lock,pdm.lock,package-lock.json,
npm-shrinkwrap.json, and piprequirements*.txtandconstraints*.txt. ACR-DEP-001(high): certifi pinned before 2026.10.18, which still trusts emSign ECC Root CA - G3, whose
intermediates' keys were compromised (GHSA-6q48-6c7w-9gw2), in any of those formats. On KServe it found two test-resource locks a manual sweep had missed.ACR-DEP-002(high): packages fetched over plain HTTP, or--trusted-host, in a lock or pip file.ACR-DEP-003:--extra-index-urlin a pip file, which lets a public package that takes an internal name win.ACR-DEP-004(low): an npm lock entry installed from git or a GitHub tarball, which has no integrity hash.ACR-APP-019: Go code that compares, measures or sets an ECDSA private key'sDwithbig.Int, with
PrivateKey.BytesandParseRawPrivateKeyas the fix.ACR-APP-020(low): Go HTTP requests with no timeout:http.Getand friends, and the default or a bare client in
files that never give a request a context or deadline. On llm-d's simulator it found a dataset download with none.ACR-APP-021(low): Pythonrequestsandurlopencalls withouttimeout=, skipping docstring examples, other
objects'urlopenmethods and tests. Ten hits across MLflow and KServe, all real calls.ACR-GHA-017(high): a checkout that keeps its credentials followed by an artifact of the workspace with hidden
files, which carries the job's token in.git/config(Unit 42's ArtiPACKED). No false positives in 249 workflows
from Argo CD, MLflow, KServe, Kubeflow and others.ACR-K8S-020: RBAC escalation paths short of*, from the Kubernetes good-practices page.escalate,bind,
impersonate,nodes/proxy, minting service-account tokens and approving CSRs are high; cluster-wide Secret reads
and admission-webhook writes, which operators such as KServe, Argo CD and Istio hold routinely, are medium. Over
the RBAC in KServe, Argo CD, KAI, Kubeflow Pipelines and llm-d it found 77 grants, 73 of them those two kinds.
Changed
- The README's rule table names the dependency area, and its rule count is tested against the registry.
v0.5.0: GitLab and JUnit reports, --diff-file, expiring suppressions, 103 rules
Added
--format gitlab(GitLab Code Quality) and--format junit(Jenkins, Azure DevOps, CircleCI, Bitbucket), plus a GitLab CI template and integration guides indocs/ci-integrations.md.--diff-file PATH|-reviews any unified diff, from a file or stdin.--min-severity/min_severity:noise floor, applied before the AI step.- Expiring suppressions:
acr:ignore[ID] until=YYYY-MM-DD; after the date the finding returns with a note. acr baseline --check(fails when accepted findings were fixed) andacr baseline --prune(removes them, never adds new ones).- AI cost and privacy controls:
--ai-max-findings(default 50, most severe first) and--ai-dry-run(prints the redacted request without calling a model). --timings, and rule/scanner durations in the JSON report;acr rules --category.- Docker Compose files and coding-agent instruction files (
AGENTS.md,CLAUDE.md,.cursorrules, Copilot instructions) are classified; the skill rules now also check instruction files. - Sixteen rules (103 in total): four Compose rules (privileged/host namespaces, Docker socket and host mounts, unpinned images, databases published on all interfaces),
curl | bashin workflows, unpinneddocker://actions, unauthenticated Lambda function URLs, purgeable Key Vaults and secrets, credentials in ConfigMaps, allow-all NetworkPolicies,enableAllProjectMcpServers, agent access to the whole disk or home directory, regex built from request input, DjangoALLOWED_HOSTSwildcard, CSRF disabled, SSH servers in images. - GitHub Action
min-severityinput; CI checks that the demo fails the gate and every report format renders. docs/architecture.md; SECURITY.md now lists exactly what the AI step sends.
Fixed
- The diff parser marked a hunk's whole range as changed, so diffs with context lines (a plain
git diffpassed to--diff-file) also reported unchanged lines. Only added lines count now. - Sixteen rules from 0.1.0 had no direct tests; each now has positive and negative cases, and a test fails if any rule is untested.
v0.4.0: custom rules, --include, agent secret redaction, SARIF fingerprints
Added
- Custom rules in
.acr.yml(custom_rules: id, title, regex pattern, severity, file globs, fix), validated on load, run like built-in rules and listed byacr rules/acr explain. --includeandinclude:to review only part of a repository;acr classifyshows a file's kinds and the rules that apply.- Nineteen rules (87 in total): hostPort, unmasked /proc, ingress-nginx snippets, privileged Pod Security namespaces, unauthenticated services (Ollama, Ray, Jupyter, Redis, ...) exposed outside the cluster, ACTIONS_ALLOW_UNSECURE_COMMANDS, toJSON(secrets), credential files copied into images, resource policies open to any principal, KMS rotation off, CloudTrail logging off, per-object bucket ACLs, model servers with --trust-remote-code, public Gradio share links, skills that pre-approve unrestricted Bash, open redirects, path traversal, deprecated TLS versions, insecure session cookies.
- SARIF: stable
partialFingerprints, tool execution notifications for failed or missing scanners and AI errors, and full rule help for built-in rules. - JSON report:
schema_versionand agatesection (threshold, blocking count, passed), plusfiles_skippedandscanners_failed. - GitHub Action:
scannersinput andfindings,blocking,report-fileoutputs. - Markdown summary links each location to the PR head commit and collapses findings beyond the first 25.
- CI fails on Python SyntaxWarnings.
Changed
- AI agent: credential files are never read or searched, and secrets in the diff and tool output are redacted before they reach the model; findings now include their category and the rule's fix guidance.
- Inline PR comments already posted by an earlier run are not repeated; GitHub API calls retry on 5xx, rate limits and network errors; review bodies stay under GitHub's size limit.
- A scanner that times out or crashes is reported as failed instead of looking clean.
- ACR-GHA-001 also covers issue_comment and workflow_run workflows that check out PR code; ACR-AI-003 covers LangChain
load_toolswith shell or REPL tools; ACR-APP-004 covers joblib, pandas.read_pickle and numpy allow_pickle; ACR-MCP-002 covers keys in server URLs; ACR-K8S-003 reports container runtime sockets as critical.
Fixed
- Files whose names git quotes (non-ASCII, quotes, backslashes) were silently skipped in every mode.
- ACR-TF-001 counted an open egress block against a restricted ingress rule, and missed all-traffic ingress (
protocol = "-1"). - hostPath written in YAML flow style was not detected.
- Shallow clones without a merge base made
--basefail; acr now falls back to a direct diff and says so.
v0.3.0: scoped suppressions, --staged, GitHub annotations, acr init
Added
- Rule-scoped inline suppressions:
acr:ignore[ACR-APP-004, semgrep:*]silences only the listed rules; a bareacr:ignorestill silences the line. acr review --stagedreviews the index (what will be committed), not unstaged edits.--format githubprints workflow-command annotations; the GitHub Action shows them by default (annotationsinput), which also works for pull requests from forks.acr initwrites a commented, validated starter.acr.yml.acr rules --format markdown|json; a test keeps the README rule catalogue in sync with the code.- Ten rules (68 in total): coding agents launched with permission checks or the sandbox off (Claude Code, Codex, Gemini CLI), agent settings that pre-approve any or risky shell commands, auto-approved MCP tools,
secrets: inheritto external reusable workflows, self-hosted runners on pull-request workflows, TLS or signature checks disabled in image builds, seccomp/AppArmor unconfined, instance metadata without session tokens (IMDSv1, GKE node metadata), predictable randomness for secrets, XML external entities. - Shell scripts, Makefiles and committed coding-agent settings files are classified, so rules can target them.
- CI installs the built wheel in a clean virtualenv and smoke-tests the CLI.
Changed
- Inline suppression is applied once, centrally, to every finding: structural rules (checked at the line they report) and external scanners now honour it too.
- The pre-commit hook runs
--stagedat the pre-commit stage only. - Severity overrides in
.acr.ymlaccept globs; an exact rule id wins, then the longest matching pattern.
Fixed
- A marker scoped to one rule no longer hides other rules' findings on the same line, and a marked
FROM ... AS stageline keeps its stage alias. - Container findings pointed at line 1 when the container's
- name:line had a trailing comment. - Plural token settings (
http_tokens = "required",max_tokens) are no longer reported as hardcoded secrets.