Back to skill

Security audit

Skill Stuff

Security checks for vulnerabilities and agentic risk

Overview

This package needs Review because a fun daily recap skill is bundled with many unrelated high-impact skill documents and the main recap also schedules broad recurring access to connected accounts.

Install only after the publisher either ships a clean main-character-recap package or clearly labels this as a multi-skill bundle with explicit permissions. Before enabling the recap cron, choose exactly which sources it may read, keep delivery private, and review or disable persistent history and memory retention.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
price-watch-SKILL.md:42
Finding

Unrestricted Product URL Fetching Permits SSRF-Style Access

Content
View full analysis
" }, "sessionTarget": "isolated", "wakeMode": "now", "payload": { "kind": "agentTurn", "message": "Run price-watch. Read {baseDir}/watchlist.md and {baseDir}/config.md. For each active item: fetch current price using web_fetch on stored URL or web_search for product. Compare to target price and last price. Alert if: price at or below target, price dropped 5%+ since last check, stock low, or sale detected. Update watchlist.md with new prices. Exit silently if no alerts.", "lightContext": true } } ``` ### Technical Analysis The skill accepts an arbitrary URL, persists it in `watchlist.md`, and directs `web_fetch` to retrieve it during both initial registration and subsequent scheduled checks. No controls are specified for: - URL scheme validation. - Loopback, private, link-local, multicast, or reserved IP ranges. - Cloud instance metadata endpoints. - Embedded credentials in URLs. - DNS rebinding. - Redirects from an approved public address to a private address. - Restriction to known product or commerce domains. If `web_fetch` has network a ...[truncated 1488 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
parav-connector-SKILL.md:183
Finding

Contradictory Parav Routing Policy Exposes Authenticated Social Sessions to Remote Infrastructure

Content
View full analysis
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:127
Finding

Daily Recap Collects Data from All Connected Sources Without Per-Source Minimization

Content
View full analysis
", "tz": "" }, "sessionTarget": "isolated", "wakeMode": "now", "payload": { "kind": "agentTurn", "message": "Run the main-character-recap skill. Read {baseDir}/shows/active.md for style. Read {baseDir}/memory.md for season context. Pull last 24h from all sources. Follow the output format in active.md exactly. After generating, update memory.md and write today's recap to {baseDir}/history/YYYY-MM-DD.md.", "lightContext": true }, "delivery": { "mode": "announce", "channel": "", "to": "", "bestEffort": true } } ``` The runtime instructions repeat the broad collection behavior: ```md ### 3. Pull sources Pull last 24h from all connected sources (calendar, email, Slack, GitHub, etc.) Privacy rules: first names only, no addresses or financial figures. If nothing returns: ask for a one-sentence description. Never fabricate. ``` ### Technical Analysis The recreational recap function does not inherently require complete access to every connected service. The instruction “Pull last 24h from all sources” establishes blanket recurring collection rather than specifying: - Which sources the user explicitly selected. - Which accounts, workspaces, channels, repositories, or folders are in scope. - Which fields may be read. - Whether message bodies are necessary. - Whether third-party confidential da ...[truncated 1780 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:373
Finding

Unpinned Community Skill-Pack Installation Creates a Mutable Supply-Chain Boundary

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (367)

Description-Behavior Mismatch

Critical
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file is a complete bill-monitoring and financial-alerting skill, while the supplied metadata says the skill is a TV-show-style daily recap tool. This is a severe capability/identity mismatch that can hide sensitive data collection, inbox access, financial tracking, and outbound actions behind an unrelated and less suspicious skill identity.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Critical
Category
Not specified by scanner
Confidence
99% confidence
Finding

The manifest/context says this is a harmless TV-show-style recap skill, but the file actually defines a job-search management workflow with persistent storage and web access. This mismatch can cause the skill to be invoked or trusted under false pretenses, exposing sensitive career data and enabling behaviors the user did not knowingly authorize.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file describes a broad thought-leadership content production and publishing system that is materially different from the manifest’s stated purpose of generating fun TV-show-style daily recaps. This kind of scope mismatch is dangerous because it can cause an agent or reviewer to grant a low-risk recap skill access to higher-risk content creation, image generation, analytics, and publishing behaviors that were not transparently declared.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The documented approval, deployment, and auto-post workflow exceeds the stated recap function and introduces outward-facing actions on third-party platforms. Hidden or undisclosed deployment capability is dangerous because users may invoke what appears to be a harmless recap tool while actually enabling publication workflows with reputational and privacy consequences.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

Auto-posting and deployment to external channels are not justified by a fun recap skill and create a direct path from generated content to public publication. In this context, that is especially dangerous because the skill’s benign branding lowers user suspicion while enabling actions that may be irreversible, reputation-damaging, or disclose personal information.

Content

No source excerpt is available for this finding.

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · SKILL.md (reported line 866)May include surrounding context.

md
- `/mcr add show "[name]" --preview` — preview before saving
- `/mcr cancel` — remove cron entirely (confirm first)
- `/mcr memory` — show current memory.md
- `/mcr reset memory` — wipe memory, start fresh season

---

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The title and documentation explicitly identify the file as 'Bill Monitor', directly contradicting the declared 'main-character-recap' skill context. This kind of overt mismatch is a strong indicator of deceptive packaging, which can be used to smuggle privileged behaviors into environments that would not approve them if accurately labeled.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill directs the agent to scan Gmail for billing emails and extract financial information, which is a materially sensitive capability. In the provided package context, this file is masquerading as a different skill entirely, so the email-access behavior is especially dangerous because it expands access to inbox data under false pretenses and could expose private financial records.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The privacy section asserts that data never leaves the local workspace and is never sent to external services, but the skill also directs the agent to access Gmail to scan receipt emails. This contradiction can mislead users about where their data is processed and undermines informed consent for highly sensitive financial and email data handling.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The setup flow tells the agent to scan Gmail for healthcare emails, pharmacy receipts, and possible test results without a clear upfront warning that this accesses highly sensitive data. Even though the document later discusses privacy rules, that does not substitute for explicit notice and consent at the moment of collection, making unauthorized or poorly understood access more likely.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill metadata and body implement inbox/email triage capabilities even though the declared skill context is a TV-show-style daily recap. This mismatch is dangerous because it can cause a user or orchestrator to grant Gmail access and run high-privilege email actions under false pretenses, enabling covert data access or action execution.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The file requests and operationalizes sensitive capabilities including reading private email, drafting replies, filing messages, scheduling recurring runs, and optionally sending autonomous responses. In the context of a purported entertainment/recap skill, these privileges are unjustified and create a strong risk of unauthorized surveillance, data exfiltration, or user-impacting actions against the inbox.

Content

No source excerpt is available for this finding.

Scope Creep

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly instructs the agent to read and update local files such as config.md and state.md even though the manifest only declares browser and web_fetch tools. This creates a capability/permission mismatch: an orchestrator or user may rely on the manifest for trust decisions, while the skill behavior expects broader filesystem access and persistent state mutation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill performs persistent management of profile, applications, and salary preferences, which is materially more sensitive than a recap feature and outside the claimed scope. Scope deception is dangerous here because job-search records can reveal intent to leave an employer, compensation targets, and application history if the skill is auto-invoked or broadly accessible.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
94% confidence
Finding

Web-based company and salary research expands the skill's capability beyond the stated recap purpose and introduces external-content handling risks, including prompt injection from job postings or company pages. In a mislabeled skill, users and platform policy may not expect or adequately sandbox these networked behaviors.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

Interview prep, follow-up tracking, and weekly pipeline review are substantial workflow features unrelated to a recap-only skill and can continuously process sensitive personal and professional information. The danger is increased by the privacy-sensitive context: these routines could surface confidential job-search activity in the wrong session or be triggered under an inaccurate trust model.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill metadata presented to the analyzer says this is a 'main-character-recap' entertainment skill, but the actual file defines a different 'meeting-notes' skill that processes meeting content, tracks commitments, and can integrate with transcripts, calendar, and email context. This mismatch is dangerous because it can bypass user/admin expectations and review controls, causing a user to invoke or install a skill with materially different data access and behavior than advertised.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest declares only web_search as an allowed tool, but the skill operationally instructs access to calendar, email, Slack, GitHub, task systems, and weather. This creates a capability-transparency failure: users, reviewers, and enforcement layers may believe the skill is narrowly scoped while it is designed to drive broad data access and aggregation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The document explicitly says social-platform posting such as LinkedIn, Reddit, Instagram, and Twitter must never be routed through parav, but later lists a skill configuration where content-publisher posts to LinkedIn and Reddit via the cloud browser. This contradiction can cause operators or downstream skills to send authenticated social actions through remote infrastructure against stated safety boundaries, increasing risk of account compromise, session misuse, and policy-violating automation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The privacy rules say the skill must not run in group or shared contexts, but the cron delivery is configured as announced channel delivery with a channel and recipient placeholder and no private-only enforcement. That creates a direct path for sensitive third-party names, relationship notes, and contact timing to be disclosed into a shared space if misconfigured or invoked in the wrong context.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

Automatic Gmail searching is a high-sensitivity capability because it grants access to private communications and can reveal far more than renewal dates. For a reminder skill, this is only justified if it is clearly in scope, narrowly constrained, and consented to; otherwise it creates unnecessary exposure of personal and financial information.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill proposes searching Gmail for renewal-related messages without a prominent, user-facing warning about access to privacy-sensitive email data. Because emails may contain identity documents, financial records, travel details, and account information, hidden or weak disclosure materially increases privacy and consent risk.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
94% confidence
Finding

The recurring /renew scan command enables repeated email rescans, which materially expands the skill from reminder management into ongoing inbox monitoring. Repeated access increases the chance of collecting unrelated sensitive data and normalizes broad surveillance without strong user awareness.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The recurring Gmail re-scan command lacks an explicit warning that each invocation may revisit private mailbox content and extract updated information. Without clear disclosure, users may trigger repeated access without understanding the privacy implications, especially in a life-admin skill handling passports, insurance, and other sensitive categories.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The manifest declares only web_search and web_fetch, but the body instructs browser automation, persistent storage, cron scheduling, and calendar modification. This mismatch hides the skill's real authority from reviewers and users, undermining consent and allowing a much broader operational scope than advertised.

Content

No source excerpt is available for this finding.