Back to skill

Security audit

verdikta-discover

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed read-only planner for scoping public, non-sensitive outside work and does not itself buy, upload, sign, or spend anything.

Use this only for public, non-sensitive tasks you are comfortable turning into an external work specification. Keep it in a no-wallet, no-secrets agent where possible, review the generated assessment before sharing it, and treat any later funding or commissioning as a separate owner-authorized action.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (33)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

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

md
1. Public `https` pages only, and screen each URL first with `scripts/url-screen.mjs` or by hand: no credentials, port, IP address, internal hostname or link sh

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 34)May include surrounding context.

md
n also run commands belongs in a sandbox that cannot reach secrets or wallets (`references/install.md`).

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 38)May include surrounding context.

md
into, the skill directory: on some hosts it is read-only, and the check script `scripts/preview.bundle.mjs` is self-contained and needs nothing installed (`scri

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 38)May include surrounding context.

md
pt `scripts/preview.bundle.mjs` is self-contained and needs nothing installed (`scripts/preview.mjs` is the unbundled source for repository checkouts); a scratc

Instruction Override

High
Category
Prompt Injection
Confidence
80% confidence
Finding

This pattern attempts to override system instructions or ignore safety constraints. Without LLM analysis, manual review is recommended.

Content

Scanner excerpt · _meta.json (reported line 39)May include surrounding context.

json
"securityNotes": [
    "No wallet, private key, seed phrase, API key, account or payment method is needed or read, and the skill never signs, funds or spends. SKILL.md forbids invoking a transactional skill or hand-writing API or RPC calls. A draft is not a quote or a purchase: amounts stay null and the result is DRAFT_NOT_QUOTED.",
    "Content the agent reads from the web is untrusted. SKILL.md has the agent treat page text as data, report injection attempts to the owner, and stop using a source whose final URL changes host. These rules reduce injection risk; they do not remove it. Where possible run the skill in an agent with no wallet, no transactional skill and no secrets, as references/install.md describes.",
    "scripts/url-screen.mjs contains prompt-injection detection patterns: regular expressions that match phrases such as 'ignore previous instructions' or 'reveal your system prompt'. They are detectors used to warn about manipulated pages, not instructions to the agent.",
    "scripts/preview.bundle.mjs is generated by scripts/build-bundle.mjs (esbuild, unminified) from scripts/preview.mjs. It inlines the skill's templates and the npm packages listed with their licenses in scripts/preview.bundle.NOTICES.txt (AJV and its helpers, and @noble/hashes). AJV compiles the skill's own fixed JSON Schemas with new Function at run time; no schema comes from input or the network. The URLs inside the bundle are JSON Schema identifiers and source comments, and nothing is fetched.",
    "references/install.md recommends that the owner add a short pointer to the agent's always-loaded instructions so the agent opens this skill before checking linked sources. The owner applies it by hand; the skill never edits agent configuration, instructions or memory.",
    "The pilot accepts only public, non-sensitive material: confidential, private, internal or regulated inputs are classified UNSUITABLE even with the owner's approval. Assessment inputs and drafts are meant to be shar
...[truncated 23 chars]

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · _meta.json (reported line 39)May include surrounding context.

json
"securityNotes": [
    "No wallet, private key, seed phrase, API key, account or payment method is needed or read, and the skill never signs, funds or spends. SKILL.md forbids invoking a transactional skill or hand-writing API or RPC calls. A draft is not a quote or a purchase: amounts stay null and the result is DRAFT_NOT_QUOTED.",
    "Content the agent reads from the web is untrusted. SKILL.md has the agent treat page text as data, report injection attempts to the owner, and stop using a source whose final URL changes host. These rules reduce injection risk; they do not remove it. Where possible run the skill in an agent with no wallet, no transactional skill and no secrets, as references/install.md describes.",
    "scripts/url-screen.mjs contains prompt-injection detection patterns: regular expressions that match phrases such as 'ignore previous instructions' or 'reveal your system prompt'. They are detectors used to warn about manipulated pages, not instructions to the agent.",
    "scripts/preview.bundle.mjs is generated by scripts/build-bundle.mjs (esbuild, unminified) from scripts/preview.mjs. It inlines the skill's templates and the npm packages listed with their licenses in scripts/preview.bundle.NOTICES.txt (AJV and its helpers, and @noble/hashes). AJV compiles the skill's own fixed JSON Schemas with new Function at run time; no schema comes from input or the network. The URLs inside the bundle are JSON Schema identifiers and source comments, and nothing is fetched.",
    "references/install.md recommends that the owner add a short pointer to the agent's always-loaded instructions so the agent opens this skill before checking linked sources. The owner applies it by hand; the skill never edits agent configuration, instructions or memory.",
    "The pilot accepts only public, non-sensitive material: confidential, private, internal or regulated inputs are classified UNSUITABLE even with the owner's approval. Assessment inputs and drafts are meant to be shar
...[truncated 23 chars]

YARA rule 'agent_skill_prompt_injection_hidden_instructions': Prompt injection or hidden instructions embedded in AI agent skill text [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · _meta.json (reported line 39)May include surrounding context.

json
njection attempts to the owner, and stop using a source whose final URL changes host. These rules reduce injection risk; they do not remove it. Where possible run the skill in an agent with no wallet, no transactional skill and no secrets, as references/install.md describes.",
    "scripts/url-screen.mjs contains prompt-injection detection patterns: regular expressions that match phrases such as 'ignore previous instructions' or 'reveal your system prompt'. They are detectors used to warn about manipulated pages, not instructions to the agent.",
    "scripts/preview.bundle.mjs is generated by scripts/build-bundle.mjs (esbuild, unminified) from scripts/preview.mjs. It inlines the skill's templates and the npm packages listed with their licenses in scripts/preview.bundle.NOTICES.txt (AJV and its helpers, and @noble/hashes). AJV compiles the skill's own fixed JSON Schemas with new Function at run time; no schema comes from input or the network. The URLs inside the bundle are JSON Schema i

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/check-result.bundle.mjs (reported line 1887)May include surrounding context.

js
function validateAsync() {
        const ruleErrs = gen2.let("ruleErrs", null);
        gen2.try(() => assignValid((0, codegen_1._)`await `), (e) => gen2.assign(valid, false).if((0, codegen_1._)`${e} instanceof ${it.ValidationError}`, () => gen2.assign(ruleErrs, (0, codegen_1._)`${e}.errors`), () => gen2.throw(e)));
        return ruleErrs;
      }
      function validateSync() {
        const validateErrs = (0, codegen_1._)`${validateRef}.errors`;

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/preview.bundle.mjs (reported line 1887)May include surrounding context.

js
function validateAsync() {
        const ruleErrs = gen2.let("ruleErrs", null);
        gen2.try(() => assignValid((0, codegen_1._)`await `), (e) => gen2.assign(valid, false).if((0, codegen_1._)`${e} instanceof ${it.ValidationError}`, () => gen2.assign(ruleErrs, (0, codegen_1._)`${e}.errors`), () => gen2.throw(e)));
        return ruleErrs;
      }
      function validateSync() {
        const validateErrs = (0, codegen_1._)`${validateRef}.errors`;

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · scripts/url-screen.mjs (reported line 22)May include surrounding context.

js
// A long token that mixes upper case, lower case and digits looks like base64 or a random key; a slug of lowercase words does not.
const looksEncoded = token => token.length >= 40 && /^[A-Za-z0-9+/_=-]+$/.test(token) && /[a-z]/.test(token) && /[A-Z]/.test(token) && /[0-9]/.test(token);
const hasEncodedSegment = decoded => decoded.split(/[/?&;=]/).some(looksEncoded);
const SECRET_FILES = /(?:^|[/\\])(?:\.env(?!\.(?:example|sample|template)\b)|id_rsa|id_ed25519|[^/]*\.pem|[^/]*keystore[^/]*|wallet\.json|\.ssh|\.aws|\.npmrc|[^/]*-bot\.json)(?:$|[/\\.?])/i;
const SECRET_PARAMS = /(?:^|[?&;])(?:password|passwd|pwd|secret|token|api[_-]?key|apikey|key|auth|authorization|signature|sig|session|cookie|mnemonic|seed)=/i;

// A bare 40-hex path segment is a git commit on a code host, not a payload. Anywhere else, or at 64 hex, it stays suspect.

YARA rule 'agent_skill_prompt_injection_hidden_instructions': Prompt injection or hidden instructions embedded in AI agent skill text [agent_skills]

High
Category
YARA Match
Confidence
80% confidence
Finding

YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).

Content

Scanner excerpt · scripts/url-screen.mjs (reported line 116)May include surrounding context.

js
s)|home director(?:y|ies)|current director(?:y|ies)|the agent'?s|this machine|the host)`;
const CATEGORIES = [
  { code: 'OVERRIDE', weight: 3, negatable: true, patterns: [
    /\b(?:ignore|disregard|forget|override|bypass|discard)\b[^.\n]{0,50}\b(?:previous|prior|above|earlier|all|any|every|your|the|these|those)\b[^.\n]{0,30}\b(?:instructions?|rules?|prompts?|guidelines?|directives?|constraints?|system prompt|programming|training|safety)\b/i,
    /\b(?:ignore|disregard|forget)\b[^.\n]{0,30}\b(?:your|the)\b[^.\n]{0,20}\b(?:owner|user|operator|principal)(?:'s|s')?\b/i,
    /\byour (?:task|instructions?|mission|goal) (?:has|have) (?:changed|been (?:updated|replaced))\b/i,
    /\b(?:forget|drop|abandon|stop) (?:about )?(?:the |your )?(?:task|job|work) (?:you were given|you are doing|at hand)\b|\bstop what you are doing\b/i,
    /\b(?:your|the (?:current|assigned|original|fact-?checking|verification|review)) task (?:is|has been) (?:suspended|cancell?ed|void|terminated|over)\b|\byour (?:new

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
89% confidence
Finding

The skill text explicitly references shell-based helper scripts such as scripts/url-screen.mjs, node scripts/screen.mjs, and preview scripts, but the skill does not declare an explicit tool scope or allowed-tools boundary. That mismatch can cause hosts or agents to assume shell use is acceptable without clear restriction, increasing the risk of command execution in contexts where shell access is broader than intended.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill description contains very broad activation triggers such as 'leaves outside help to your judgment' and allows use when outsourcing is merely worthwhile, which can cause the skill to activate on ordinary tasks the agent could complete itself. In this context, over-triggering is risky because the skill is specifically about drafting external work/handoffs, so mistaken activation can push data-sharing or delegation workflows unnecessarily close to execution.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
81% confidence
Finding

This JSON request describes a task in general natural language but does not define a narrow invocation trigger, exclusion condition, or explicit scope boundary. In a manifest-like file, broad operational instructions without trigger specificity can increase the risk of unintended activation or misuse in systems that map descriptions to skill selection.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file asserts a concrete, physically performed visit in its step log and attestation while also declaring itself a synthetic demonstration. In a skill designed to structure outsourced real-world work and evidence packs, this contradiction can normalize fabricated evidence formats or mislead downstream consumers into treating mock results as genuine field verification.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · references/install.md (reported line 10)May include surrounding context.

hermes skills install verdikta/verdikta-applications/skills/hermes/verdikta-discover

text

Obtain the complete [skill directory](https://github.com/verdikta/verdikta-applications/tree/main/skills/verdikta-discover) and [entrypoint](https://raw.githubusercontent.com/verdikta/verdikta-applications/refs/heads/main/skills/verdikta-discover/SKILL.md) from main. Pin the commit you reviewed; branch URLs are mutable. The ClawHub release (`clawhub install verdikta-discover`) has the same files without `tests/`; the test and evaluation files this page names are in the repository. Copy the complete directory, or install the release, to the host runtime’s skill location and follow that runtime’s loader instructions. Native loading has been verified on OpenClaw 2026.8.33 (see `tests/EVALUATION_PROTOCOL.md`). On Hermes Agent the skill needs its own copy; it has been verified there with that copy for claim checks against linked sources (Hermes rounds 1 and 2) and, with the pointer's second part, for the review and real-world templates (Hermes rounds 3b to 3e; section "Hermes Agent" below; `tests/CONNECTED_DESIGN.md`).

Host configuration. Primary posture: an agent with web fetch that follows the hard rules in `SKILL.md`, with no signer, no transactional skill, no wallet environment and read-only file tools. Reading task sources needs no owner approval step; every URL is screened first with `scripts/url-screen.mjs` (also `node scripts/screen.mjs url <url>`). The screens are pure functions: `screenUrl` (https only; no credentials, port, IP literal, internal hostname or shortener; no query string, secret-shaped value or task text in a URL the agent composed), `screenRedirect` (a final origin that differs from the requested one makes the source unavailable), `screenContent` (advisory heuristics for text that tries to instruct the agent) and `isPublicIp` (for a host to check what a name resolved to). They reduce risk and do not make a page trustworthy: a URL screen cannot see an injecti
...[truncated 25 chars]

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% 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 · references/install.md (reported line 14)May include surrounding context.

md
Host configuration. Primary posture: an agent with web fetch that follows the hard rules in `SKILL.md`, with no signer, no transactional skill, no wallet environment and read-only file tools. Reading task sources needs no owner approval step; every URL is screened first with `scripts/url-screen.mjs` (also `node scripts/screen.mjs url <url>`). The screens are pure functions: `screenUrl` (https only; no credentials, port, IP literal, internal hostname or shortener; no query string, secret-shaped value or task text in a URL the agent composed), `screenRedirect` (a final origin that differs from the requested one makes the source unavailable), `screenContent` (advisory heuristics for text that tries to instruct the agent) and `isPublicIp` (for a host to check what a name resolved to). They reduce risk and do not make a page trustworthy: a URL screen cannot see an injection inside a reputable page, and the content screen misses much of what a human would catch (see `tests/EVALUATION_PROTOCOL.md` for measured rates). The limits on what the agent can do (no secrets, no spending tools, no uploads) are the real control.

Where the runtime has a pre-call hook, run `screenUrl` there so a blocked URL is never fetched. OpenClaw's documentation describes a `before_tool_call` plugin hook that can return `{ block: true, blockReason }` for a `web_fetch` call, fail-closed and scoped per agent, and a built-in guard that blocks private and internal addresses. That hook has not been run on a live gateway for this skill, and OpenClaw's documentation says it cannot rewrite `web_fetch` results, so content screening stays advisory there. The safest option remains a dedicated discovery agent with no web fetch at all: on OpenClaw 2026.8 use `tools.allow: ["read"]` and a `skills` allowlist of `verdikta-discover` (example: `tests/openclaw/agent.patch.example.json`).

Executable preview: from this skill directory run `node scripts/preview.bundle.mjs examples/assessment.json`, or pipe the input t
...[truncated 26 chars]

Vague Triggers

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

This JSON file is a manifest/result-style file, so SQP-1 applies. The file contains structured result data but no explicit invocation conditions, trigger phrases, or exclusion conditions, leaving trigger scope unspecified if this file is used to describe skill activation behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

This JSON describes a skill-related preview artifact but does not define any explicit trigger phrases, activation conditions, or exclusion conditions. For manifest-scope review, that absence leaves invocation scope ambiguous and could contribute to unintended use if this file is relied on to describe when the skill should run.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

This JSON file describes a task result and contains no explicit activation conditions, trigger phrases, or scope constraints for when the associated skill should run. For manifest-style files, the absence of any trigger specificity can make invocation boundaries unclear if this file is used operationally to define or describe the skill.

Content

No source excerpt is available for this finding.

Known Vulnerable Dependency: esbuild==0.27.4 — 1 advisory(ies): GHSA-g7r4-m6w7-qqqr (esbuild allows arbitrary file read when running the development server on Window)

Low
Category
Supply Chain
Confidence
89% confidence
Finding

The lockfile pins esbuild 0.27.4, which is reported as affected by GHSA-g7r4-m6w7-qqqr: arbitrary file read via esbuild's development server on Windows. This is a real dependency risk, although exposure is limited here because esbuild is a devDependency and the issue is only relevant if the development server feature is used in a vulnerable Windows development environment.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.

Content

Scanner excerpt · package.json (reported line 12)May include surrounding context.

json
"bundle": "node scripts/build-bundle.mjs"
  },
  "dependencies": {
    "@noble/hashes": "^1.8.0",
    "ajv": "^8.17.1",
    "ajv-formats": "^3.0.1"
  },

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.

Content

Scanner excerpt · package.json (reported line 13)May include surrounding context.

json
},
  "dependencies": {
    "@noble/hashes": "^1.8.0",
    "ajv": "^8.17.1",
    "ajv-formats": "^3.0.1"
  },
  "devDependencies": {

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
40% confidence
Finding

Dependencies lack version pinning, allowing potential malicious package updates. Consider pinning versions.

Content

Scanner excerpt · package.json (reported line 14)May include surrounding context.

json
"dependencies": {
    "@noble/hashes": "^1.8.0",
    "ajv": "^8.17.1",
    "ajv-formats": "^3.0.1"
  },
  "devDependencies": {
    "esbuild": "0.27.4"

Known Vulnerable Dependency: esbuild==0.27.4 — 1 advisory(ies): GHSA-g7r4-m6w7-qqqr (esbuild allows arbitrary file read when running the development server on Window)

Low
Category
Supply Chain
Confidence
60% confidence
Finding

Dependency has known vulnerabilities (CVEs). Using packages with unpatched security flaws exposes the environment to known exploits.

Content

No source excerpt is available for this finding.