Back to skill

Security audit

Cite Holmes — Deep Research × Hallucination-Free Citations

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because it reports serious weaknesses in citation verification and a network redirect protection flaw; the scan is partial, so its evidence is incomplete.

Treat the suspicious verdict as a reason to review carefully before installing. The strongest concerns are whether citation results can be trusted and whether network requests respect private-network and proxy boundaries. The scan covered 98.8% of components but was partial, including a parser limit and a missing referenced path, so it cannot establish that unreported behavior is safe.

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

Warning
Location
scripts/ch_netio.py
Finding

Hostname-based redirects bypass private-network request protection

Content
View full analysis

Impact

An attacker can cause internal HTTP HEAD requests and, on the implemented fallback paths, GET requests. Reachability and HTTP status are reflected in verification results, providing an internal-service probing oracle; requests can also invoke internal endpoints that act on these methods. Arbitrary response-body disclosure, cloud credential theft, and code execution are not established.

Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (120)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied code's primary function is skill-package linting, not citation authenticity or literature research. Its checks concern SKILL.md fields and publishing compatibility; it performs no literature searches, registry queries, citation verification, retraction checks, or bibliography exports. Although this could be a maintenance utility within a larger research skill, the supplied chunk implements a separate packaging-validation purpose rather than a supporting step in the declared research workflow. It only reads a local file and prints results; no network access, file modification, or malicious behavior is evident.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

test_arxiv_api_down_page_200_still_verified expects verified when the arXiv metadata API fails and the landing page merely returns HTTP 200. Likewise, test_s2_mismatch_silent_no_downgrade and test_s2_404_silent_skip expect verified after DOI metadata retrieval fails and Semantic Scholar provides mismatching metadata or no record, provided the URL is reachable. Reachability does not authenticate the submitted title, year, or other citation details, so these expectations contradict the description's full-verification guarantee. This conclusion concerns behavior explicitly expected by the tests; the imported verify_refs implementation was not supplied. Network lookups, temporary files, exports, and security-hardening tests otherwise support the declared purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The only supplied executable behavior is npx skills add docsor1212/cite-holmes, an installation command. It does not itself perform the declared research and verification workflow. Although this may be legitimate setup documentation for the advertised skill, the installed skill's implementation is absent and its claimed capabilities cannot be established from this chunk.

Content

No source excerpt is available for this finding.

Parallel deduplication substitutes a previously verified reference for a distinct citation sharing its identifier.

High
Category
Excessive Agency
Confidence
98% confidence
Finding

Parallel followers inherit the leader's entire verification result when their identifier-based cache keys match, without independently validating each follower's citation metadata. Rebuilding semantic_audit does not check a changed title, author, or year, so an altered citation can inherit a genuine reference's verified verdict and even disappear behind the leader's copied metadata, undermining this skill's central authenticity guarantee.

Content

No source excerpt is available for this finding.

Cross-language verification can accept an OpenAlex title match without establishing that it belongs to the cited DOI.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The OpenAlex bridge searches the claimed title and retrieves only display_name, never establishing a relationship between the matching work and the cited DOI. A real title from one work paired with an unrelated DOI in another writing system can therefore produce a successful bridge and bypass the DOI-title mismatch check, undermining the skill's core authenticity guarantee when this keyed fallback is enabled.

Content

No source excerpt is available for this finding.

References can receive a verified verdict without successful citation-authenticity verification.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The final verdict gate requires only a trusted tier and complete fields, not successful metadata verification. Failed registry checks can therefore fall through to verified; accepting a caller-supplied tier further weakens this gate, allowing fabricated citation details to acquire an authoritative verdict.

Content

No source excerpt is available for this finding.

Cloud Metadata Access

High
Category
Server-Side Request Forgery
Confidence
90% confidence
Finding

Code accesses a cloud instance metadata endpoint (e.g. 169.254.169.254). A single request can return temporary IAM credentials, making this a high-value SSRF target for credential theft.

Content

Scanner excerpt · references/security.md (reported line 24)May include surrounding context.

md
**已实施的防护**:
- 入口字面量防护:URL 主机部分为回环/内网/链路本地/保留/多播 IP 字面量
  (`127.0.0.1`、`169.254.169.254` 云元数据端点、`10/8`、`192.168/16` 等)
  或 `localhost`/`*.local`/`*.internal` 主机名时,拒绝外呼并如实标注
  (判定落 unreachable,不冒充任何注册库结论)。
- 重定向逐跳复核(v6.6.0):自动跟随 30x 的路径上,每一跳目标重新过同一

Cloud Metadata Access

High
Category
Server-Side Request Forgery
Confidence
90% confidence
Finding

Code accesses a cloud instance metadata endpoint (e.g. 169.254.169.254). A single request can return temporary IAM credentials, making this a high-value SSRF target for credential theft.

Content

Scanner excerpt · tests/test_v60600.py (reported line 55)May include surrounding context.

python
def test_entry_guard_semantics(self):
        # 字面量防护语义(既有 ssrf_blocked 的回归面收拢);测试数据的
        # 私有网段字面量拼接构造(随包泄漏扫描不过测试数据面,v6.5.0 教训)
        for bad in ("http://127.0.0.1/x", "http://169.254.169.254/meta",
                    "http://10." + "1.2.3/x",
                    "http://192." + "168.5.5/x",
                    "http://localhost:9/x", "http://svc.internal/x"):

Batch deduplication reuses another reference's verification result instead of checking each reference's claims.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The parallel deduplication path copies the leader's complete verification result to followers without independently checking their claimed metadata; rebuilding semantic_audit does not revalidate those claims. An attacker can place a legitimate reference before a fabricated reference sharing its identifier and cause the latter to inherit a verified verdict, undermining this skill's central citation-authenticity guarantee.

Content

No source excerpt is available for this finding.

OpenAlex title-only search can validate a title paired with an unrelated DOI.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The OpenAlex bridge searches only the claimed title and requests only display_name, so it cannot establish that the matching work has the DOI being verified. When this fallback is reached and an OpenAlex key is configured, an existing title paired with an unrelated DOI can produce a successful bridge and potentially be reported as verified. This undermines the skill's central protection against stitched citations, especially in medical research.

Content

No source excerpt is available for this finding.

References can receive a verified verdict despite unsuccessful registry metadata verification.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The final verdict logic can mark a complete reference from a trusted tier as verified without requiring successful registry metadata verification. Network failures or skipped checks can therefore produce an authenticity endorsement, undermining the skill's central guarantee and potentially laundering fabricated or mismatched references into research or medical outputs.

Content

No source excerpt is available for this finding.

The regression test requires a verified verdict despite failed official metadata verification and a mismatching fallback record.

High
Category
Excessive Agency
Confidence
96% confidence
Finding

The regression test explicitly requires a verified verdict when official metadata retrieval fails, the fallback record contradicts the claimed citation, and only landing-page reachability succeeds. This codifies a fail-open authenticity policy that can launder fabricated or misattributed references into verified output, with heightened consequences in the skill's medical research context; the production implementation is not included here for independent confirmation.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The documented npx skills add command can download and execute an unpinned release of the skills CLI, and the skill repository is also not fixed to an immutable revision. This creates a supply-chain exposure if an upstream release is compromised; it is an installation command, however, not an MCP server reference as the finding claims.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

This repeated installation instruction executes the skills CLI without an exact version and retrieves mutable skill content. A compromised upstream release could execute with the installing user's privileges, although the command is not itself an MCP server reference.

Content

No source excerpt is available for this finding.

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

Reference-verification workflows lack a privacy disclosure about external reference lookups.

Medium
Category
Output Handling
Confidence
85% confidence
Finding

The integration guidance directs agents to submit their reference lists for verification without explaining that lookups disclose bibliographic identifiers or search terms to external services. Although external lookups serve the stated purpose, automatically processing references from confidential drafts can reveal research interests without an informed privacy decision.

Content

No source excerpt is available for this finding.

The onboarding workflow directs users to verify their own references without disclosing external data transmission.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The onboarding path moves directly from supplying a personal reference list to running verification without disclosing external lookup traffic. References from unpublished or proprietary work can reveal sensitive research topics even when the complete document is not transmitted.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'uvx from'.

Medium
Category
MCP Rug Pull
Confidence
98% confidence
Finding

The recommended uvx installation and MCP configurations execute code from a mutable Git repository without selecting an immutable revision. A compromised upstream repository or malicious update could consequently execute with the agent host's permissions; this is a supply-chain exposure, not evidence of an existing compromise.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'uvx from'.

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The local-checkout installation workflow clones the repository's default branch and then runs the package without checking out a reviewed revision. Although the local uvx invocation does not itself select a remote version, its preceding clone introduces the same mutable-upstream execution risk.

Content

No source excerpt is available for this finding.

The documented offline mode still permits outbound preflight connections.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The preflight branch makes outbound socket connections without checking args.offline. Combining --offline with --preflight therefore violates the advertised no-network boundary and exposes connection metadata, although these probes do not transmit the bibliography itself.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · mcp/ch_registry.py (reported line 86)May include surrounding context.

python
try:
            q = quote(claimed.strip(), safe="")
            url = (
                f"https://api.openalex.org/works?search=%7Bq%7D&per-page=1"
                f"&select=display_name&api_key={quote(_ROOT._OPTS['openalex_key'], safe='')}"
            )
            if _ROOT._cb_open(url):

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · scripts/ch_registry.py (reported line 86)May include surrounding context.

python
try:
            q = quote(claimed.strip(), safe="")
            url = (
                f"https://api.openalex.org/works?search=%7Bq%7D&per-page=1"
                f"&select=display_name&api_key={quote(_ROOT._OPTS['openalex_key'], safe='')}"
            )
            if _ROOT._cb_open(url):

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · mcp/ch_registry.py (reported line 182)May include surrounding context.

python
# v3.6 --cn 回源:api.crossref.org/works/{doi} 返回同构 CSL(message 字段)。
        # 独立主机不计入 doi.org 断路器;失败静默回原路径。
        try:
            cr_url = f"https://api.crossref.org/works/{quote(doi, safe='()/')}"
            cr_req = urllib.request.Request(
                cr_url, headers={"User-Agent": ua_contact, "Accept": "application/json"}
            )

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · mcp/ch_registry.py (reported line 1648)May include surrounding context.

python
"check_failed": False,
        }
    mailto = _ROOT._OPTS.get("mailto") or ""
    url = f"https://api.crossref.org/works/{quote(doi, safe='()/')}"
    if mailto:
        url += f"?mailto={quote(mailto, safe='@')}"
    if _ROOT._cb_open(url):