Back to skill

Security audit

Codebase Audit: audit

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious because it found concerns about broad activation, approval handling, and potentially unsafe operations; some referenced content was only partially inspected.

Treat this as a review item before installing. Consider whether you trust its source-changing workflow and whether you can limit its access to the files and shell it needs; do not rely on approvals stored in a handoff file as proof of authorization. Coverage was partial: 21 of 22 components were scanned, and one shell script was only partially inspected.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/scan.py
Finding

Repository-controlled paths become executable shell syntax in analyzer commands

Content
View full analysis

Impact

An attacker can execute commands with the auditing agent's filesystem and environment access, enabling unauthorized file modification or access to credentials.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (20)

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 21)May include surrounding context.

md
- `scripts/verify.py`: checks findings files against the lens files. `scripts/cochange.sh`: files that change together.

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
80% confidence
Finding

Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.

Content

Scanner excerpt · lenses/design.md (reported line 14)May include surrounding context.

md
han one place (client and server enums, a docs table and the code it describes, two validators). Change one and the others must change too, but nothing says so. Fix family: one source of truth that the others derive from or are generated from.
- **global-state**: A module reads a global, singleton, environment value or file that another module sets. Change that module and this one breaks, but nothing shows it.
- **scattered-feature**: One feature split across folders by layer, or related code in different folders, so one change needs all of them and the layout hides it. `cochange.sh` pairs with no import between them are the candidates. Fix family: colocate by domain.
- **implicit-protocol**: A call order or convention callers just have to know (init before use, call `track()` after every action). Change it and callers break with no warning. Fix family: make it explicit in the interface or impossible to skip.

**Needless links**: pieces tied together that change for different reasons.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · DELIVER.md (reported line 26)May include surrounding context.

md
Delivery charts the report onto the repo's issue tracker as a map plus child tickets, the way `wayfinder` charts an effort. The map is an **index, not a store**: each decision or fix lives in exactly one ticket, and the map gists it and links.

Where it goes: the tracker in `docs/agents/issue-tracker*.md` when `setup-matt-pocock-skills` has configured one; otherwise the local-markdown tracker below. Publishing to a remote tracker (GitHub, GitLab, Linear) posts for others to see, so confirm with the user first; the local tracker needs no confirmation.

Local-markdown layout:

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The resume procedure treats a writable HANDOFF.md approval field as sufficient authority to modify source without renewed confirmation. Anyone able to alter or supply that artifact can forge approval or redirect an approved ticket, turning persisted workflow data into authorization.

Content

Scanner excerpt · SKILL.md (reported line 30)May include surrounding context.

md
arrive (CI, `claude -p`), unanswered questions stay on `HANDOFF.md` as open questions, delivery goes to the local tracker only, and source stays untouched unless a ticket is on the `approved:` line.
- Create the output directory with a `.gitignore` containing `*`, so the audit touches no tracked file. When starting a new run (never when resuming), first move an earlier run's `RUN.md`, `REPORT.md` and `findings/` to `<out>/history/<its date>/`, so the new run can compare against it.
- **Resume**: if `<out>/HANDOFF.md` exists, continue from its `next` phase, leaving the run's files where they are. Tickets on its `approved:` line are cleared for source changes without asking again; work the first one on the frontier, one ticket per session, as `DELIVER.md` says. Any other ticket goes back to the user first. If an earlier run exists in `history/`, offer `--since <its commit>` and carry `refuted.md` forward.

Every phase below ends by rewriting `<out>/HANDOFF.md` as `DELIVER.md` describes.

The skeleton-generation instruction lacks a warning to preserve original source files.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

The instruction to script away function bodies does not specify generating a separate representation rather than rewriting working source. In an agent-driven refactor, that ambiguity can cause destructive edits, including loss of uncommitted implementations that Git cannot recover.

Content

No source excerpt is available for this finding.

The unconditional activation rule does not limit the skill to relevant code-review tasks.

Medium
Category
Trigger Abuse
Confidence
90% confidence
Finding

"Fires: always" specifies no invocation context or exclusions. It can activate the failure-analysis workflow for unrelated requests rather than only reviews involving application failure handling or concurrency.

Content

No source excerpt is available for this finding.

The process-termination test lacks a warning or isolation requirement to protect user data.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The workflow explicitly instructs the agent to kill a process between stages of a multi-step write. This can interrupt real writes and leave local or remote data inconsistent, but the skill provides no warning or requirement to use an isolated test environment.

Content

No source excerpt is available for this finding.

“Fires: always” specifies unconditional activation without limiting the performance lens to relevant tasks.

Medium
Category
Trigger Abuse
Confidence
91% confidence
Finding

The activation condition applies universally, without restricting invocation to performance analysis or code-review tasks. This can cause the skill to activate for unrelated requests; the conditional checks later in the file limit individual checks, not activation of the lens itself.

Content

No source excerpt is available for this finding.

The unconditional activation condition has no task or context boundaries.

Medium
Category
Trigger Abuse
Confidence
93% confidence
Finding

“Fires: always” activates the security review regardless of whether the user's task involves software or security. The file provides no activation exclusions, so unrelated requests could trigger this entry-point and secret-rotation workflow.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · lenses/state.md (reported line 11)May include surrounding context.

md
least-once; exactly-once does not exist. For each write, is the second run harmless? Fix family: idempotency key plus a unique constraint, or a state transition that is a no-op when repeated.
- **out-of-order** (when clients sync offline or concurrent writes can interleave): Offline mutation B arrives before A; an old response lands after a new one. Does an old write win? Look for last-write-wins by arrival time, server `now()` stamped at replay, no version or sequence column, responses applied without checking they are still current.
- **calendar-adversary** (when the code computes dates, times, durations or schedules): Run each date computation at 23:59 and 00:00, on a DST change, on Feb 29, on Dec 31 (ISO week 53 vs calendar year), after the user flies east and west (Tel Aviv → LA), with a non-Gregorian device calendar, and in 12h and 24h locales. Expiry dates, streaks, weekly quotas and "today" live here.
- **truth-owner**: Find every fact stored twice: cache and DB, client enum an

The activation condition “Fires: always” provides no scope or exclusions, allowing the test-strength workflow to activate on unrelated tasks.

Medium
Category
Trigger Abuse
Confidence
99% confidence
Finding

The skill explicitly activates unconditionally rather than limiting invocation to test-strength audits or mutation-testing requests. Its workflow includes running tests and modifying code, so applying it outside the intended testing context could cause unintended work.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/scan.py (reported line 171)May include surrounding context.

python
def git(root, *args):
    try:
        return subprocess.run(["git", "-C", root, *args], capture_output=True, text=True, check=True).stdout
    except (subprocess.CalledProcessError, FileNotFoundError):
        return None

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · scripts/signals.tsv (reported line 4)May include surrounding context.

text
type	id	lenses	scope	pattern	tag
# stack rows name the platform; signal rows fire lenses; seed rows pre-locate candidate sites for lens.probe; special rows only count
# scope: path = regex on file path | dep = regex on manifest contents | code = all source files | ext,ext = those extensions | test: prefix = test files only
stack	ios	evolution	path	\.xcodeproj/|(^|/)Info\.plist$|\.entitlements$	-
stack	swift	-	path	(^|/)Package\.swift$|\.swift$	-
stack	android	evolution	path	(^|/)AndroidManifest\.xml$	-
stack	jvm	-	path	(^|/)(build\.gradle(\.kts)?|pom\.xml)$	-

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · scripts/signals.tsv (reported line 46)May include surrounding context.

text
seed	data.user-b	-	code,sql,rules	(?i)row level security|using \(true\)|security definer|service_role|allow read, write|req\.body\.user_?[iI]d|params\.user_?[iI]d	-
seed	data.migration-on-real-data	-	sql	(?i)add constraint|set not null|create (unique )?index (?!concurrently)|alter column .* type	-
seed	data.delete-account	-	code,sql	(?i)delete_?account|deleteUser|delete_user|on delete (cascade|set null)|auth\.admin\.deleteUser	-
seed	security.secrets	-	code,json,yaml,yml,plist,xml	(?i)(api[_-]?key|secret|token|password)\s*[:=]\s*['\"][A-Za-z0-9_\-]{16,}['\"]|sk_live_[0-9a-zA-Z]{10,}|AKIA[0-9A-Z]{16}|-----BEGIN (RSA |EC |OPENSSH )?PRIVATE KEY	-
seed	security.injection	-	code	execute\(f['\"]|execute\([^)]*\+|\.raw\(|\$queryRawUnsafe|innerHTML\s*=|dangerouslySetInnerHTML|\beval\(|shell=True|Runtime\.getRuntime\(\)\.exec|child_process|NSPredicate\(format: ?\"[^\"]*\\\(	-
seed	security.local-storage	-	code	(?i)(UserDefaults|localStorage\.setItem|SharedPreferences|AsyncStorage\.setItem)[^\n]

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · scripts/signals.tsv (reported line 49)May include surrounding context.

text
ve_[0-9a-zA-Z]{10,}|AKIA[0-9A-Z]{16}|-----BEGIN (RSA |EC |OPENSSH )?PRIVATE KEY	-
seed	security.injection	-	code	execute\(f['\"]|execute\([^)]*\+|\.raw\(|\$queryRawUnsafe|innerHTML\s*=|dangerouslySetInnerHTML|\beval\(|shell=True|Runtime\.getRuntime\(\)\.exec|child_process|NSPredicate\(format: ?\"[^\"]*\\\(	-
seed	security.local-storage	-	code	(?i)(UserDefaults|localStorage\.setItem|SharedPreferences|AsyncStorage\.setItem)[^\n]*(token|password|jwt|secret|session)	-
seed	security.transport	-	code,plist,xml	http://(?%21localhost%7C127%5C.0%5C.0%5C.1%7Cschemas%7Cwww%5C.w3%5C.org%29%7CNSAllowsArbitraryLoads%7CusesCleartextTraffic%7Cverify%5Cs%2A=%5Cs%2AFalse%7CrejectUnauthorized%3A%5Cs%2Afalse%7CInsecureSkipVerify	-
seed	security.entry-points	-	code,xml,plist	onOpenURL|open url:|CFBundleURLSchemes|intent-filter|WKWebView|WebView\(|addJavascriptInterface|javaScriptEnabled|postMessage\(|multipart/form-data	-
seed	security.confused-deputy	-	code	(fetch|axios\.get|requests\.get|URLSession\.shared\.data)\(\s*(req|request|params|body|url)\

Session Persistence

Medium
Category
Rogue Agent
Confidence
75% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · scripts/signals.tsv (reported line 50)May include surrounding context.

text
\(|shell=True|Runtime\.getRuntime\(\)\.exec|child_process|NSPredicate\(format: ?\"[^\"]*\\\(	-
seed	security.local-storage	-	code	(?i)(UserDefaults|localStorage\.setItem|SharedPreferences|AsyncStorage\.setItem)[^\n]*(token|password|jwt|secret|session)	-
seed	security.transport	-	code,plist,xml	http://(?%21localhost%7C127%5C.0%5C.0%5C.1%7Cschemas%7Cwww%5C.w3%5C.org%29%7CNSAllowsArbitraryLoads%7CusesCleartextTraffic%7Cverify%5Cs%2A=%5Cs%2AFalse%7CrejectUnauthorized%3A%5Cs%2Afalse%7CInsecureSkipVerify	-
seed	security.entry-points	-	code,xml,plist	onOpenURL|open url:|CFBundleURLSchemes|intent-filter|WKWebView|WebView\(|addJavascriptInterface|javaScriptEnabled|postMessage\(|multipart/form-data	-
seed	security.confused-deputy	-	code	(fetch|axios\.get|requests\.get|URLSession\.shared\.data)\(\s*(req|request|params|body|url)\b|createSignedUrl\(|service_?[rR]ole	-
seed	security.supply-chain	-	json	\"[\w@/.-]+\"\s*:\s*\"(\^|~|\*|latest)|\"postinstall\"	-
seed	security.privacy-flows	-	code	Analytics\.|analytics\.|\.track\(|logEvent\(|Sentr

The unconditional activation rule does not limit this design-review lens to relevant tasks.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

“Fires: always” activates the lens without restricting it to code design or structural review. No exclusion conditions or invocation context are provided, so unrelated tasks could trigger unnecessary repository analysis.

Content

No source excerpt is available for this finding.

The activation condition “Fires: always” does not constrain invocation to stateful-code or time-related reviews.

Low
Category
Trigger Abuse
Confidence
93% confidence
Finding

The skill declares unconditional activation without specifying a surrounding review workflow or excluding unrelated requests. This could invoke its diagramming and analysis requirements for tasks that contain no relevant code or stateful entities.

Content

No source excerpt is available for this finding.

The unconditional activation rule does not limit this lens to type-modeling reviews.

Low
Category
Trigger Abuse
Confidence
95% confidence
Finding

“Fires: always” specifies activation without any task or context restriction. Although the remaining content concerns type modeling, the activation rule does not exclude unrelated requests and could cause unintended invocations.

Content

No source excerpt is available for this finding.