ast02-supply-chain-compromise
Skill2 starsChanged 3 months agoApache-2.0View on GitHub
Detect and triage OWASP AST02 Supply Chain Compromise — registry flooding, dependency confusion in a skill's nested requirements.txt/package.json, config-file hijacking (.claude/settings.json, hooks), and maintainer-account takeover. Use when auditing a skill registry's provenance controls, when a skill's top-level files look clean but its dependency tree is unaudited, when a repository's config files (not its source) are the suspected execution path, or when deciding whether a finding belongs here versus AST01 (the payload) or AST08 (the missed detection).
- Reads credentials
In this repository10 other files it gives its agents.
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 29
Stop now: if the route below sends you elsewhere.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AST02 - Supply Chain Compromise
Unofficial community implementation of the OWASP® Agentic Skills Top 10 standard, v1.0. NOT an OWASP project; no OWASP endorsement. OWASP is a registered trademark of the OWASP Foundation, used here descriptively to identify the standard implemented. Source: https://owasp.org/www-project-agentic-skills-top-10/
Pattern: Knowledge. Read the orientation block, take a route, then descend only as far as
the route sends you. The one mechanism a package can decide — an execution path in a config
file the host auto-reads at project open — lives in scripts/detector.py; the frozen
scenario tiers, and the three scenarios no package decides, live in coverage-matrix.md.
Orientation — read this much, then decide whether to read the rest
- Fires when the question is delivery and provenance: how a skill got onto the machine, whether the thing that arrived is the thing the publisher signed, and whether a config file the host opens unprompted carries an execution path.
- Decides exactly one scenario in-package —
AST02-S03Config-File Hijacking. One check, deliberately. The rest of the file is the audit reasoning for the controls a reviewer has to check by hand. - Never decides Registry Flooding, Dependency Confusion or Maintainer Account Takeover. Those are properties of a corpus, a resolver and an account; none is inside the package you are holding, and the emptiness of the matrix is the measurement.
- Stop now if the route below sends you elsewhere. This is the category where the in-artifact surface is smallest, so a wrong route wastes the most reading.
Freedom, stated up front, because one narrow half of this category is mechanical and the rest of it is an audit:
| Half | Freedom | Why |
| The one in-package scenario | Low — the keying decides, you do not | AST02-S03 fires on the config surface, not on how alarming a command reads: the scanned path list is closed, the auto-executed shapes are four, and the command keys are a fixed set. Widening any of those is an edit to scripts/detector.py that somebody reviews, never a call made while reading a result. What tier a named scenario holds is settled above this file — scenarios/registry.yaml is rank 2 of coverage-matrix.md's authority chain and scripts/detector.py is rank 5, subordinate to it. Where a row in Route first and a neighbor's own Distinguishing section disagree, the neighbor fixed that seam and this file is the bug. |
| Everything registry-side | High — there is no check to defer to | Registry Flooding, Dependency Confusion and Maintainer Account Takeover are audited by a person or not at all, and the decision rules below hand you tests rather than verdicts: which of the three provenance questions an implementation actually performs, whether a pin reaches every package in the transitive tree, whether revocation is addressable at the granularity a given incident needs. The evidence for each lives in a corpus, a resolver and an account, so no rule can close them from here. The one thing that is not free: a judgement you reach this way stays a judgement — it does not get filed under a scenario id. |
Route first
| If the finding is | Go |
| a payload's content — what the delivered skill actually does | AST01 — stop reading here |
| a poisoned document the skill merely reads at runtime, which no hash pin reaches | AST05 — stop reading here |
| an agent auto-applying an upstream update without human review, or an unpatched install | AST07 — stop reading here |
| a scanner that never walks the recursive dependency tree | AST08 — stop reading here |
| provenance, signing, revocation, dependency pinning, or a repo config file that executes on open | this skill — continue below |
| a Maintainer Account Takeover incident | both — the takeover is AST02, the blind auto-apply is AST07; read the neighbors section, not the whole file |
| unclear, or it reads like several of these at once | node cli/bin/cli.js route "<finding text>" names the primary category and lists contributing control failures separately |
Then descend only this far
| Section | Read it when |
| Why this category has almost no in-artifact surface | you are about to claim registry membership as a provenance signal, or to ask why only one check ships |
| Decision rules | always: six rules, and the reason this file exists |
| Distinguishing AST02 from its neighbors | the route above was contested, or one incident spans two categories |
| What the one shipped check decides, and the six ways it goes quiet | MANDATORY before you report a negative AST02-S03 result. Every one of the six returns a negative verdict that is not a clean one |
| NEVER | always, and again before the verdict is written down. The conclusions the silences above and the empty three-quarters of the matrix invite; each names the check, scenario id or test that refutes it |
| Scope and out-of-artifact boundary | you are being asked to detect a maintainer-account compromise from the artifact |
Do NOT load coverage-matrix.md to decide whether something is a finding. It is the
tier contract and the F1 denominator; open it only to cite the authoritative tier of a
named scenario, or the written reason a scenario is out-of-artifact. Do NOT load
scripts/detector.py to learn what the one check decides — the quiet list below states
its keying, its four firing shapes and its six silences in prose. Read the source only to
extend the closed path list.
Why this category has almost no in-artifact surface
Three of AST02's four named scenarios are properties of a corpus over time, not of a package: Registry Flooding is a publication rate, Dependency Confusion is a resolver outcome in a namespace the package does not contain, and Maintainer Account Takeover produces a release that verifies exactly as an honest one does — the attacker holds the legitimate key. One package is indistinguishable from one member of a flood, and the registry records no in-package signal for any of the three. That is why this skill ships exactly one check and why its coverage matrix is mostly empty: the emptiness is the measurement, and padding it with proxy checks filed under scenario ids would be an overclaim, not coverage.
The corollary for review practice: "the skill is on a registry" carries near-zero provenance signal. Registry membership and cryptographic verification are two different assertions answered by two different mechanisms; a lookup that returns found has not verified anything.
Decision rules
Split by the freedom table above, so that the voice is not mistaken for uniform. Two are
closed and act on bytes you are holding: rule 5 is the reasoning the shipped check
implements, and rule 2's in-package half is already decided elsewhere — a declared
content_hash contradicting the shipped bytes is AST01's check, recorded as a category
precondition. The other four are audit instruments. They tell you what to ask of a
registry, a resolver, a dependency tree and a revocation endpoint; each leaves the verdict
with you, because the thing that would settle it is not in the package.
- "Listed", "signed", and "unrewritten" are three separate questions, and a system that answers only the first reports PASS for all three. A membership lookup answers does this digest appear in what the log currently serves. It does not answer is this the digest the publisher actually signed (that needs a signature over a canonical serialization, verified against a resolvable publisher key), and it does not answer has the log's own history been rewritten (that needs inclusion and consistency proofs against a monitored root). The failure mode is silent by construction: each unanswered question degrades to a pass, so the composite verdict looks identical whether all three controls exist or only the cheapest one does. When auditing, name which of the three a given implementation performs — the absence of the other two is the finding, and it will not appear as an error anywhere.
- Sign the bundle, not the entry point. The signature must cover a canonical
digest of SKILL.md plus every declared resource file, so a post-publish edit to
any declared file — not just the top-level manifest — invalidates it. A signature
scheme that only hashes SKILL.md leaves every referenced resource file an
unauthenticated attack surface. The in-package half of this is decidable and this
repository decides it: a declared
content_hashthat contradicts the shipped bytes is a self-contained contradiction, owned by AST01's check and recorded there as a category precondition — never as coverage of an AST02 scenario. - The payload is usually in the nested dependency, not the top-level skill.
Dependency Confusion tampers with a transitive package, not the surface skill file
— this is precisely why a scanner that only inspects the skill's own top-level
files (the AST08 failure mode) misses it. A skill named "Summarize YouTube Videos"
that imports
yutube-dl-coreinstead of the legitimate package is the canonical shape: the surface skill reads clean. - Version ranges are not a security control; hash pins are.
requests>=2.25.0permits any future release under that range, including a compromised one, to satisfy the dependency at install time with no new review.requests==2.31.0 --hash=sha256:<digest>pins to bytes, not to a mutable name+range. Hash-checking mode requires every package in the transitive tree to carry a hash — a partial pin is not a pin. - Repository configuration files are executable code, not passive metadata — treat
them as first-class attack surface.
.claude/settings.json, hooks, and environment-override files (e.g.ANTHROPIC_BASE_URL) can trigger on repo-open, before any user action and before any skill even runs — this is a distinct trigger condition from AST01's "user installs/invokes a malicious skill." - Revocation must be addressable at three granularities, or it isn't usable. A compromised signing key, one bad skill version by content digest, and an entire publisher are three different blast radii; a revocation mechanism that can only revoke "the whole publisher" over-blocks every other skill from a large publisher for one bad release, and under-scopes when only a single key is compromised. Hosts must consult a revocation endpoint at load time with a bounded freshness window — a revocation list checked only at publish time misses post-publish compromise.
Distinguishing AST02 from its neighbors
- vs AST01: AST02 is delivery/provenance; AST01 is payload. "Publishers could upload unlimited packages with no scanning" is an AST02 finding about the registry even if every uploaded skill that day happened to be benign.
- vs AST05 (Untrusted External Instructions): AST02's code-integrity controls (hash pins, signed digests) can pin and verify a dependency. They cannot pin a document a skill merely reads at runtime — that is AST05's surface exactly because those integrity controls do not reach it. A CVE fixed in a pinned dependency and a poisoned referenced runbook are different attack surfaces requiring different controls, even when both are "external content the skill trusts."
- vs AST07 (Update Drift): AST02 is compromise at publish/delivery time (Maintainer Account Takeover pushing v2.0 with a payload); AST07 is drift after a legitimate install goes unpatched, or an update mechanism blindly applying upstream changes. The Maintainer Account Takeover scenario sits at the seam: the takeover itself is AST02, and an agent auto-applying that malicious update without human review is the AST07 half of the same incident.
- vs AST08: a scanner that never inspects the recursive dependency tree — only the top-level skill file — is an AST08 gap that causes AST02 dependency-confusion attacks to go undetected. Classify the missed detection capability as AST08 and the underlying compromise as AST02; they are not interchangeable labels for one finding.
What the one shipped check decides, and the six ways it goes quiet
AST02-S03 is keyed on the config surface, not on the presence of a command
string: the same shell command is unremarkable inside a bundled script and is an
execution path the host enters unprompted inside .claude/settings.json. CVE-2025-59536
and CVE-2026-21852 are why the keying is that way round — cloning and opening a
repository was sufficient, with no dialog shown. Within those files the check fires on
four auto-executed shapes only: a hook entry carrying a command, an MCP server entry
that spawns a process, a control-plane environment override, and a folder-open task.
MANDATORY before you report a negative — the reading is not the duty, the writing is. A negative here is a two-part deliverable: the verdict, and which of the six silences below the package in front of you is actually sitting inside, named from that package's own files rather than recited as a list. A reader handed a negative and no boundary supplies the boundary themselves, and the one they supply is everything — so an unqualified negative asserts a coverage this check has never had.
A positive is the same duty pointed the other way: quote the path the check matched, the key it hung off inside that file, and the command string it read. The check is keyed on the config surface rather than on how the command reads, so the surface is the finding's identity — a report that names the category and paraphrases the command sends its reader looking for something to remove and gives them no entry to remove it from. Each of the following returns a negative verdict that is not a clean one.
- The check reads config files shipped inside the package. The incident shape is a config file in the repository the agent opens. That repository is not part of any skill package, so a clean AST02-S03 result on a skill says nothing about the workspace it will run in. Scan the target repo separately; this check does not.
- The scanned path list is closed and host-specific. Eight exact path tails are
covered.
.devcontainer/devcontainer.json(whosepostCreateCommandis precisely this scenario),.envrc,.github/workflows/*,.zed/,.idea/, and any newer agent host's config directory are not. Extend the list per host before relying on a negative. - The parse is strict JSON, and an unparseable file ends the whole scan. A settings
file with comments or trailing commas (the JSONC dialect editors actually accept), or
a YAML/TOML equivalent, raises — and the check returns immediately with
unparseable JSON; no execution path decided, so config files sorted after it are never examined. That is an INCOMPLETE result wearing a negative verdict; read the evidence string, not the boolean. - Environment overrides are only found inside a block named
env,environment, orenvVars. VS Code's dottedterminal.integrated.env.osxkey and any override set at the document root fall outside that shape and are missed even when the variable itself is on the control-plane list. - An MCP command must hang off a named server child. A command scalar placed
directly on the
mcpServersmapping is skipped deliberately, so the container itself cannot be convicted — a host that reads a command from that position is uncovered. - Command detection is a fixed key set of scalars (
command,cmd,script,exec,run,shellcmd). An entry that carries onlyargs, or that spells the keyentrypointorprogram, yields no command value and clears — inside a hook block that the host will still execute.
NEVER
The check above is narrow on purpose and three of four scenarios are out of reach, so this category's characteristic error is not a missed pattern — it is a conclusion drawn from a silence. Each entry names what refutes it, in this directory.
- NEVER close the config-hijacking question for a deployment on the strength of a clean
package scan.
AST02-S03's trigger in CVE-2025-59536 and CVE-2026-21852 is a cloned repository the agent opens; the check's entire input is the package's own shipped files. The one delivery path the scenario is named after is the one the check structurally cannot see, so a negative here is filed as though the incident shape had been examined when it was never in scope. Scan the workspace as a separate target. - NEVER read a negative
AST02-config-file-hijackingas "every config file was examined". The first file that fails to parse returns immediately, so a.claude/settings.jsoncarrying a JSONC trailing comma ends the run before.vscode/tasks.jsonis opened — and the Finding it returns isdetected=False. Any aggregator that reads the boolean records the category clean; only the evidence string saysunparseable JSON; no execution path decided. - NEVER add a host's config path to the closed list without a parser for that host's
dialect.
.envrcis shell,.github/workflows/*is YAML,.devcontainer/devcontainer.jsonis JSONC. Each one added under the strict-JSON parse turns every package that ships it into the undecided verdict above and abandons the files sorted after it. Widening the surface that way makes the check quieter, not louder — the opposite of the intended change. - NEVER treat
package ships no config file a host auto-reads at project openas a statement about a host outside the eight covered path tails. A.devcontainer/devcontainer.jsonwhosepostCreateCommandis precisely this scenario returns that sentence verbatim — the same sentence a package with no config file at all gets. An uncovered host and a clean package are indistinguishable in the output, so the gap never surfaces in a report unless a reader knows the list is closed. - NEVER re-key the check on command-looking strings instead of on the config surface.
That ablation is written and scored:
tests/test_corpus_discriminates_mechanism.pyruns it over this corpus at F1 0.667 (tp 2, fp 1, fn 1) against the shipped check's 1.000. It fails in both directions at once — it misses theANTHROPIC_BASE_URLcase, which carries a URL and no command, and convicts the clean folder-open case, which carries the byte-identical command with norunOntrigger. - NEVER file a pin-posture observation under
AST02-S02. Version ranges rather thansha256:pins is that scenario's declaredartifact_signal, itsartifact_signal_checkslist is empty, and pin posture is a lockfile property while the scenario is a resolver outcome. This is not hypothetical: the pin-posture pair is one of the six orphan fixturestests/test_coverage_matrix_ast07_ast08.py::test_ast02_ships_no_orphan_fixture_corpusnames by directory so the delisting cannot be quietly undone. Declining the id does not licence inventing one in its place. Every scenario id in this suite is a row inscenarios/registry.yaml; an identifier you cannot point at there is not an identifier, however well-formed it looks, and a well-formed invention travels further than a wrong-but-real id precisely because the reader who tries to look it up finds nothing to contradict. Write the condition you observed and leave the id field empty. - NEVER report a Maintainer Account Takeover as ruled out because the release verified.
The registry's written reason for tiering
AST02-S04out-of-artifact is that a release pushed by an attacker holding the legitimate key is byte-indistinguishable from an honest one. A passing signature read as a negative converts "no in-package signal exists" into "no compromise occurred"; the honest verdict is undecidable-from-here, and it escalates to auth logs, signing-key provenance and publication cadence, never to another scan. - NEVER give a takeover incident one label to keep it on one ticket. The compromise is AST02 and the agent auto-applying the resulting update without human review is AST07. The two halves have different fixes and different owners — revocation addressable at the three granularities of decision rule 6, versus a human gate on the update path — and a single label schedules exactly one of them while closing the incident.
Scope and out-of-artifact boundary
Maintainer Account Takeover is a registry-side event — whether a given publisher
account was actually compromised is not observable from the skill artifact itself; a
detector can only observe symptoms (a signing-key change, an anomalous version jump)
that a coverage matrix may tier as agent-judgable rather than static-detectable. The
authoritative tier and its written reason live in coverage-matrix.md; this file does
not pre-empt that lock.
References — what is loadable, and when
| Resource | Load it when |
coverage-matrix.md | you must cite the authoritative tier of a named AST02 scenario, or the written reason three of the four are out-of-artifact |
scripts/detector.py | you are extending the closed config-path list for a new host — never to find out what the check decides |
| the whitepaper's own AST02 section (source: the OWASP Agentic Skills Top 10 publication, section AST02 (no local copy: this repository points at the publication rather than bundling it, so the authority chain cannot drift from a stale copy; the whitepaper is CC BY-SA 4.0)) | you need the full attack-scenario catalog, the real-world config-hijacking evidence, the Trail of Bits (Jun 3, 2026) scanner-bypass result, or the preventive-mitigation list |
This file is the delta on top of that section, not a restatement of it.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

