Back to skill

Security audit

ars-academic-paper

Security checks for vulnerabilities and agentic risk

Overview

This is a large academic writing workflow skill with expected manuscript, citation, formatting, and review behavior, but it should be used with care around private drafts and citation-checking workflows.

Before installing, expect this skill to process full drafts, bibliographies, reviewer comments, and optional prior writing samples. Review any auto-corrected citations, avoid uploading confidential or co-authored material without permission, and remember that working drafts may contain hidden audit comments until the final formatting step removes them.

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
  • 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
Findings (120)

Hidden Instructions

High
Category
Prompt Injection
Confidence
95% confidence
Finding

The skill instructs the agent to emit hidden HTML comments carrying machine-readable citation metadata that is invisible in normal rendering but preserved in output. Hidden channels are risky because downstream systems or users may not notice embedded metadata, and prompt/content boundaries can blur when invisible markers are later parsed by other agents or tools.

Content

Scanner excerpt · references/draft_writer_agent.md (reported line 416)May include surrounding context.

md
When emitting any citation in the draft body, write the citation in two layers:

1. **Visible layer**: standard author-year form (e.g. `Smith (2024)` or `(Smith, 2024)`).
2. **Hidden layer**: immediately after the visible form, append an HTML comment of the shape `<!--ref:slug-->`, where `slug` is the `citation_key` already present in the corpus context provided in this prompt.

Examples: `Smith (2024) <!--ref:smith2024-->` or `(Smith, 2024)<!--ref:smith2024-->`.

Hidden Instructions

High
Category
Prompt Injection
Confidence
97% confidence
Finding

This extends the hidden-channel pattern by embedding both reference identifiers and claim anchors in invisible comments after every citation. Because these markers are intended for downstream machine consumption, they create a covert instruction/data channel inside user-visible text that can be abused for parser confusion, hidden prompt transport, or integrity issues if any consumer treats comment contents as trusted metadata.

Content

Scanner excerpt · references/draft_writer_agent.md (reported line 435)May include surrounding context.

Every visible citation in the draft body MUST be followed by BOTH a slug marker AND an anchor marker:

text
<visible> <!--ref:slug--><!--anchor:<kind>:<value>-->

Anchor kinds (closed enum):

Hidden Instructions

High
Category
Prompt Injection
Confidence
88% confidence
Finding

This finding reinforces that the skill's logic depends on hidden inline markers as part of its control flow for temporal checks. Using invisible comment tokens as semantic triggers increases fragility and can let unreviewed hidden content influence downstream behavior or audits without clear user awareness.

Content

Scanner excerpt · references/draft_writer_agent.md (reported line 520)May include surrounding context.

md
Before writing any sentence that:

- Cites a document with a publication year via <!--ref:slug-->
- States that one event led to / was enabled by / superseded / followed another
- Uses present-tense or deictic framing ("currently", "now", "the most recent",
  "the latest", "new", "recently", "last year", "nowadays")

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/formatter_agent.md (reported line 383)May include surrounding context.

md
1. A literal `[UNVERIFIED CITATION — NO ORIGINAL]` marker (HIGH-WARN; v3.7.1).
2. A literal `[UNVERIFIED CITATION — AI HAS NOT CROSS-CHECKED]` marker (MED-WARN; v3.7.1).
3. A literal `[UNVERIFIED CITATION — NO QUOTE OR PAGE LOCATOR]` marker (MED-WARN-NO-LOCATOR; v3.7.3).
4. Any `<!--ref:slug-->` HTML comment with status neither `ok`, explicitly acknowledged plain `LOW-WARN`, nor `LOW-WARN-PARTIAL-COVERAGE` (the finalizer pass either failed or was skipped). `LOW-WARN-PARTIAL-COVERAGE` ( as superseded by ) is legal only for an active user declaration whose deterministic resolver state is `partial_coverage` or `coverage_unknown`; `not_attested` and `rescinded` remain plain unacknowledged `LOW-WARN`, and `ledger_invalid` emits `READ-LEDGER-INVALID`, which is intentionally outside this allowlist and MUST refuse. Surface the exact resolver state/reason in `provenance_summary.md`; never mislabel any non-covered state `ok`.
5. **Any `<!--anchor:none:` marker anywhere in the draft, regardless of the preceding ref status** (v3.7.3 codex round-8 F20 closure). A stale or skipped finalizer pass can leave `<!--ref:slug ok--><!--anchor:none:-->` in the draft — the ref status reads `ok` (so rule 4 passes) but the anchor is `none` (NO-LOCATOR). Since v3.7.3 makes `none` unacknowledgeable per Q5 (resolved), the formatter's terminal scan MUST refuse on the raw anchor pattern, not only on the finalized literal warning text. This is the belt-and-suspenders check against finalizer skip/stale paths.
6. A literal `[HIGH-WARN-CLAIM-NOT-SUPPORTED]` annotation (v3.8 §3.6 8-row matrix; UNSUPPORTED + source-level defect_stage). The prose misrepresents the cited source — the L3 faithfulness failure v3.8 exists to catch. Mirrors v3.7.3 R-L3-1-A asymmetry —  does NOT clear this; remediation is fixing the prose (re-cite, drop claim, or revise).
7. A literal `[HIGH-WARN-NEGATIVE-CONSTRAINT-VIOLATION` annotation (v3.8 §3.6; UNSUPPORTED + negative_constraint_violation). The author explicitly declared 
...[truncated 24 chars]

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% confidence
Finding

Skill instructs the agent to never refuse or to always comply. Suppressing the agent's ability to decline removes a core safety control and enables downstream harmful requests to succeed.

Content

Scanner excerpt · references/formatter_agent.md (reported line 417)May include surrounding context.

md
Render `finding: unresolved` or status `not_checked`, `unknown`, or `degraded` as **NOT CLEAN — UNRESOLVED**; never call those states passed, absent, or clean.
`terminal_policy.eligible: false` cannot trigger refusal or `HIGH-BLOCK`. A v1.1 eligible retraction row is still only an input to the finalizer; this formatter never evaluates it. Migration/deprecation authority is `references/shared/bibliographic_integrity_signals.md`.

For a v1.2 `tortured_phrase_match` row, transcribe `HEURISTIC-ADVISORY`, `UNMEASURED`, its exact `cited_title` or `cited_abstract` surface, counts, snapshot source/version/as-of/raw SHA-256, detached-manifest SHA-256, and any reason code. The category label is the neutral **Phrase-list screening advisory**; use **phrase-list match requiring review** only when `finding: detected`. A checked zero means only that no match was observed on that checked surface; a missing abstract stays `not_checked` / `unresolved` with `ABSTRACT_MISSING`, and a whitespace-only abstract uses `ABSTRACT_EMPTY`. These rows remain `display.marker_token: null` and `terminal_policy.eligible: false`: they never refuse conversion, alter a gate, mint or rewrite a citation marker, suggest replacement prose, or support an AI-, author-, papermill-, misconduct-, origin-, contextual-validity-, false-positive/false-negative-, precision-, recall-, coverage-, publisher-acceptance-, or clean-draft claim.

For an own-draft `tortured-phrase-advisory/1.0` artifact, accept only the already validated artifact produced over the exact conversion input and render its single bounded page of at most 25 match-detail rows. Do not re-run matching, infer context, alter counts, traverse an unbounded detail view, or edit the manuscript. Its `HEURISTIC-ADVISORY` / `UNMEASURED` result is advisory only and never a formatter refusal condition.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/formatter_agent.md (reported line 439)May include surrounding context.

md
## ARS Marker Stripping

When emitting converted **final outputs** (LaTeX / DOCX / PDF / clean-markdown deliverables), strip every ARS HTML-comment marker — `<!--ref:...-->`, `<!--anchor:...-->`, `<!--block:...-->` — from the converted artifact. One rule for all marker kinds: markers are audit metadata, not manuscript content, and must not leak into a submission (do not rely on Pandoc happening to drop raw HTML comments — strip explicitly so the rule holds across all five output formats).

Ordering is load-bearing: stripping happens ONLY AFTER every marker-dependent gate above has run against the working draft — the Cite-Time Provenance Hard Gate, the v3.10 stamp gates, and (in patch-mode revision rounds, `references/guides/revision_patch_protocol.md`) the apply script's own self-checks. The **working draft and the `phase6_*/` provenance artifacts keep their markers untouched** — `<!--block:-->` markers are the anchor layer the next revision round's manifest is generated from; stripping them there would destroy patch mode. Word counts already exclude markers regardless (`references/shared/word_count_conventions.md` § HTML-comment markers), so stripping does not change any reported count.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/guides/revision_patch_protocol.md (reported line 15)May include surrounding context.

md
| Artifact | Produced by | Convention |
|---|---|---|
| Anchored draft | `ars_anchorize_draft.py` (in place) | every block carries `<!--block:BNNNN-->`; IDs never renumbered |
| Block manifest | same run, sidecar | `<draft>.block-manifest.json` — `base_draft_hash` + `{block_id, old_hash, first_line_excerpt}` per block; the ONLY legitimate hash source for a patch |
| Immutable roadmap | editorial synthesizer / confirmed standalone adapter | `revision-roadmap/1.0`, exact draft + block-manifest bindings |
| Claim surface manifest | deterministic registry builder | `claim-surface-manifest/1.0`, exact Claim Intent + UTF-8 surface bindings |

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% 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 · references/literature_strategist_agent.md (reported line 67)May include surrounding context.

md
| `general_social_science` | Social Science **or** Policy (matches either) |
| `unknown_user_defined` | none — neutral default never triggers a mismatch check |

The mismatch advisory fires only when PCR `Discipline` falls **outside** the profile's implied set. `unknown_user_defined` and an absent row never warn.

**There is no `HANDOFF_INCOMPLETE` concern** — the profile is a PCR field, not a handoff object, so the general handoff-validation convention never applies. A profile defect simply falls back to neutral with the advisory.

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 · references/peer_reviewer_agent.md (reported line 157)May include surrounding context.

md
## Output Discipline

Keep your review **brief but complete**. State each finding and your verdict directly; do not pad them with repeated qualifiers, apologetic framing, or restated caveats. Concise does **not** mean under-caveated — preserve every material uncertainty and limitation; cut only redundancy and hedging that adds no information. One clear statement of a caveat beats three softened ones. (This governs *your own* output; it is distinct from your assessment of the paper's writing quality.)

*Epistemic status: these are prompt-surface instructions. They make the reviewer's output discipline explicit; they do not, and cannot, prove the model stays pressure-stable at runtime — that would need a separate non-deterministic behavioral eval.*

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [agent_skills]

High
Category
YARA Match
Confidence
95% confidence
Finding

This schema contains bidirectional/invisible Unicode control characters in a regex used to validate relative paths, and the YARA match also flags repeated metadata/description context consistent with tool-schema poisoning patterns. Invisible characters can mislead reviewers about what inputs are actually permitted or denied, creating a Trojan Source-style inspection gap in a security-sensitive validation contract that downstream agents may trust.

Content

Scanner excerpt · references/shared/contracts/evidence/evidence_row_v1_2.schema.json (reported line 5)May include surrounding context.

json
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://github.com/Imbad0202/academic-research-skills/shared/contracts/evidence/evidence_row_v1_2.schema.json",
  "title": "ARS Shared Evidence Row v1.2 — cross-document consistency",
  "description": "Closed observation-level evidence-row/1.2 carrier for the cross-document consistency advisory. One row integrity-binds the complete ordered bilateral or preregistration three-witness evidence set. Evidence and checked-scope states are caller-supplied semantic observations whose bytes are replayed; they are not agreement, completeness, readiness, acceptance, or authorization findings.",
  "type": "object",
  "additionalProperties": false,
  "required": [
    "schema_version",
    "surface",
    "row_id",
    "pair_kind",
    "observation_key",
    "deviation_id",
    "deviation_

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [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 · references/shared/contracts/re_review/legacy/v1_0/traceability.schema.json (reported line 5)May include surrounding context.

json
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://github.com/Imbad0202/academic-research-skills/shared/contracts/re_review/traceability.schema.json",
  "title": "Re-review Phase 2B Traceability Sidecar (machine-readable Schema 11 sidecar)",
  "description": "Phase 2B (letter-revealed) final traceability artifact of the  Spec B three-gate re-review contract. Every difference between a committed 2A verdict and the final matrix verdict rides a typed AdjustmentRecord (closed basis set, §3.4); the 2A new-issue set is carried FROZEN (whole-record byte-identical, §13 witness); all user-answerable states (§6 G2) resolve through typed records inside the deferral loop, each re-emission incrementing revision and superseding the prior emission by hash. the deterministic checker (not shipped; rules apply) recomputes every invariant and the �

YARA rule 'agent_skill_mcp_tool_poisoning_metadata': MCP/tool metadata poisoning indicators in tool schemas or skill manifests [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 · references/shared/contracts/re_review/traceability.schema.json (reported line 5)May include surrounding context.

json
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://github.com/Imbad0202/academic-research-skills/shared/contracts/re_review/traceability.schema.json",
  "title": "Re-review Phase 2B Traceability Sidecar (machine-readable Schema 11 sidecar)",
  "description": "Phase 2B (letter-revealed) final traceability artifact of the  Spec B three-gate re-review contract. Every difference between a committed 2A verdict and the final matrix verdict rides a typed AdjustmentRecord (closed basis set, §3.4); the 2A new-issue set is carried FROZEN (whole-record byte-identical, §13 witness); all user-answerable states (§6 G2) resolve through typed records inside the deferral loop, each re-emission incrementing revision and superseding the prior emission by hash. the deterministic checker (not shipped; rules apply) recomputes every invariant and the �

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
85% 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 · references/socratic_mentor_agent.md (reported line 30)May include surrounding context.

md
When the user proposes a paper RQ, thesis sentence, literature-gap statement, or chapter framing, run a light wording/framing check before continuing the normal Socratic paper-planning flow. This advisory is about **surface phrasing only**, not about idea quality, novelty, feasibility, contribution, or whether the user is "right." Same idea phrased in domain-native vocabulary should not trigger the advisory.

**Trigger rule:** compare the user's wording against the reference pattern set below. Fire only when the surface wording clearly matches one or more patterns with high confidence. If the match is weak, ambiguous, or depends on interpreting the idea content, do not warn.

**Reference phrasing patterns:**

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The README advertises very broad natural-language triggers such as 'write a paper', 'academic paper', and multilingual equivalents, which can overlap with ordinary user requests and cause the skill to activate unexpectedly. In an agent gateway, this increases the chance of unintended prompt takeover, where normal writing or review requests are routed into this skill’s multi-step workflow without explicit user consent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The metadata trigger list contains broad phrases like 'write paper', 'guide my paper', and 'parse reviews' that can match common user requests outside the intended safe routing boundaries. In agent ecosystems, overly broad activation can cause the wrong skill to engage, increasing the chance of unnecessary access to user materials or execution of a more powerful workflow than intended.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
74% confidence
Finding

Style calibration asks users to provide three or more prior papers so the system can learn their writing voice, which increases collection of sensitive authored content and may create expectations of retention or profiling beyond the immediate task. In an academic context, prior papers can contain unpublished ideas, confidential collaborations, or identifiable metadata, so this feature raises meaningful privacy and data-minimization concerns.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The file says the orchestrating session should never abort the task on a failed runtime check and should degrade gracefully instead. Later sections explicitly state that in v3.6.6 there is no graceful fallback and that writer/evaluator phase failures emit [GENERATOR-PHASE-ABORTED] and abort the full-mode run, which directly conflicts with the earlier instruction as written.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The multilingual trigger keyword blocks include many ambiguous, high-frequency phrases that could activate the skill on loosely related requests. Because this skill orchestrates a multi-phase pipeline and can request substantial user material, misrouting is more dangerous here than in a simple single-purpose helper.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instructions require the agent to produce bilingual abstracts specifically in English and Traditional Chinese, and later prescribe zh-TW-specific formatting and writing behavior. This is a natural-language locale constraint that does not provide user opt-in or explain that the skill is limited to a region-specific or user-selected context.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill hard-codes a bilingual label requirement ('[SPECULATION — UNVERIFIED] / 【推测·待验证】') for unsupported claims without indicating user opt-in or locale negotiation. This can override user-preferred language, leak unintended language/context into outputs, and create compliance/usability issues in workflows that require monolingual, jurisdiction-specific, or accessibility-constrained output.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The instruction to 'auto-correct detected errors' authorizes direct modification of user content without clearly warning that manuscript text and references may be changed. In a writing pipeline, silent edits can corrupt scholarly meaning, introduce inaccuracies, or overwrite user intent without review.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file says the sole deliverable is a Citation Compliance Report, but later sections instruct the agent to output a corrected draft and corrected reference list. Conflicting authority boundaries are dangerous because they make downstream behavior unpredictable and can cause the agent to exceed scope, modify manuscript content, or emit artifacts other phases were supposed to control.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The agent's stated role is citation-format compliance, but these instructions expand it into plagiarism adjudication and source replacement decisions. That broadening can cause the skill to make high-stakes academic-integrity judgments or alter cited support beyond its narrowly scoped mandate, increasing the chance of incorrect removals or unsupported editorial intervention.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Requiring cross-reference of all journal citations against Retraction Watch introduces external-source vetting and publication-status adjudication that exceeds a citation checker's expected scope. It can also trigger unnecessary outbound lookups about user manuscripts and lead the agent to suppress or replace sources based on incomplete or misinterpreted external data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill instructs external retraction-database checks but does not disclose that citation metadata or manuscript-derived source lists may be transmitted or compared against third-party services. That creates a privacy and confidentiality risk, especially for unpublished manuscripts, sensitive collaborations, or embargoed work.

Content

No source excerpt is available for this finding.