Security audit

Batch Cognition

Security checks for vulnerabilities and agentic risk

Overview

This skill is a batch prompt executor, but it stores complete user or Drive content indefinitely and can execute imported document text without clear confirmation boundaries.

Review before installing. Use it only for batches and Drive folders you are comfortable copying into local workspace files and reusing across sessions. Avoid secrets, credentials, regulated data, or confidential documents unless retention, deletion, redaction, and explicit confirmation rules are added.

A.I.G

Vulnerability Patterns

Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
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 (2)

T01 · Skill Instruction Hijacking

Error

Location: SKILL.md:60
Finding: Untrusted Batch and Drive Content Can Be Executed as Agent Instructions
Content: View full analysis
Remediation: View remediation

T09 · Insecure Skill Coding Practices

Warning

Location: references/rolling-decay-memory.md:45
Finding: Complete Batch Content Is Persisted Indefinitely Without Sensitive-Data Controls
Content: View full analysis
[full prompt text] ``` ``` From `SKILL.md:113-116`: ```markdown Append 🟢 items to `systems/batch-cognition/value-stack.md`. Append 🟡 items to `systems/batch-cognition/parked.md`. Append 🔴 items to `systems/batch-cognition/discarded.md`. Log any cross-batch connections to `systems/batch-cognition/connection-graph.md`. ``` From `SKILL.md:130-134`: ```markdown At each checkpoint: 1. Score all items: `salience *= 0.8`, re-referenced items → `1.0` 2. Drop items below 0.2 threshold (write tombstone, archive to disk) 3. Carry forward: rolling summary + surviving high-salience items + value stack 4. New block header written to chain file ``` From `references/rolling-decay-memory.md:45-52`: ```markdown ## Recovery Nothing is deleted. The chain is on disk. To recover: 1. Read block headers backward until you find the tombstone 2. The tombstone links to the original batch doc + item number 3. Pull the full content from the batch doc This means: **working context is always lean, but total knowledge is always complete.** ``` From `references/rolling-decay-memory.md:90-92`: ```markdown - Chain index: `systems/batch-cognition/chain/INDEX.md` (links all chains) - Token budget p ...[truncated 2939 chars]
Remediation: View remediation

SkillSpector

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Rogue AgentSelf-Modification, Session Persistence
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection

Findings (13)

Missing User Warnings

High

Category: Not specified by scanner
Confidence: 95% confidence
Finding: The skill explicitly supports Drive and file-based ingestion but gives no warning about handling potentially sensitive external documents. That combination can lead to large-scale ingestion of confidential data into persistent local memory structures without informed consent or minimization.
Content: No source excerpt is available for this finding.

Missing User Warnings

High

Category: Not specified by scanner
Confidence: 98% confidence
Finding: The skill mandates saving the entire batch to persistent files before processing, but it does not warn users that their content will be stored and replicated in local artifacts. This is dangerous because users may paste sensitive prompts, credentials, business data, or personal information expecting transient handling rather than durable storage.
Content: No source excerpt is available for this finding.

Context-Inappropriate Capability

High

Category: Not specified by scanner
Confidence: 97% confidence
Finding: The self-improvement section allows the skill to change itself based on prior runs, which is not necessary for batch processing and introduces unauthorized behavior drift. In practice, this can weaken safeguards over time, embed tainted lessons from adversarial input, and make the skill's future behavior less predictable and less auditable.
Content: No source excerpt is available for this finding.

Self-Modification

High

Category: Rogue Agent
Confidence: 99% confidence
Finding: The directive to 'Update this skill if any rule changes warranted' is explicit self-modification authority. Self-modifying skills are especially dangerous because untrusted batch inputs can indirectly influence future behavior, alter safeguards, and create persistence for adversarial instructions beyond the current session.
Content: Scanner excerpt · SKILL.md (reported line 140)May include surrounding context.
md
After each batch, append to `workspace/systems/batch-cognition/learnings.md`: - What worked / what was wasted / grouping improvements / play-think split effectiveness - Update this skill if any rule changes warranted. ## Key Rules

Vague Triggers

Medium

Category: Not specified by scanner
Confidence: 95% confidence
Finding: The activation scope covers vague patterns like bulk input, file-based lists, and Drive dumps, which can easily match ordinary user behavior and trigger the skill unexpectedly. Over-broad activation is dangerous here because the skill performs automatic saving, classification, and persistence of content once activated.
Content: No source excerpt is available for this finding.

Vague Triggers

Medium

Category: Not specified by scanner
Confidence: 86% confidence
Finding: The activation instructions provide trigger phrases but do not define meaningful negative examples or guardrails, so the skill may activate on ambiguous conversational cues. Because activation leads to data capture and execution-oriented processing, mistaken invocation can expose private content and start workflows the user did not intend.
Content: No source excerpt is available for this finding.

Ssd 3

Medium

Category: Not specified by scanner
Confidence: 95% confidence
Finding: The instruction to save the full batch and read prior value-stack/meta-think files establishes persistent reuse of complete user-provided content across sessions without sensitivity boundaries. This creates a memory corpus of user inputs that can leak confidential information into later outputs or unauthorized files.
Content: No source excerpt is available for this finding.

Ssd 3

Medium

Category: Not specified by scanner
Confidence: 94% confidence
Finding: The batch doc format retains full prompt text and adds classifications and notes under each item, multiplying the number of places where sensitive user data can be copied. This propagation increases exposure, complicates deletion, and makes accidental reuse or disclosure more likely.
Content: No source excerpt is available for this finding.

Missing User Warnings

Medium

Category: Not specified by scanner
Confidence: 84% confidence
Finding: The skill continues processing automatically after 30 seconds of no response, which treats silence as consent. In a workflow that may execute prompts, perform research, and persist content, auto-continuation can cause unintended actions and additional data handling without renewed user approval.
Content: No source excerpt is available for this finding.

Ssd 3

Medium

Category: Not specified by scanner
Confidence: 93% confidence
Finding: The value stack, parked, discarded, and connection-graph files create durable aggregation of prior user content into reusable memory artifacts. Even if each individual note seems harmless, long-term accumulation can reconstruct sensitive projects, interests, strategies, or proprietary material across batches.
Content: No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium

Category: Not specified by scanner
Confidence: 94% confidence
Finding: The skill claims to process prompt batches, but its behavior expands into persistent cross-batch memory management, value ranking, archival, and self-improvement. This scope expansion increases data retention and authority beyond what a user would reasonably expect from a batch executor, creating a hidden stateful system that can accumulate sensitive user content over time.
Content: No source excerpt is available for this finding.

Vague Triggers

Medium

Category: Not specified by scanner
Confidence: 93% confidence
Finding: The trigger condition is broad enough to activate on common user language about Google Drive or folder processing, which can cause the skill to engage unintentionally in contexts the user did not clearly consent to. Because this skill is designed for bulk processing and downstream file inspection, ambiguous activation increases the chance of over-collection and unintended handling of sensitive data.
Content: No source excerpt is available for this finding.

Missing User Warnings

Medium

Category: Not specified by scanner
Confidence: 96% confidence
Finding: The instructions tell the agent to open Google Drive, enumerate all files, and inspect file contents without any user-facing warning, scope confirmation, or minimization controls. In a Drive context, this can expose sensitive personal, business, or regulated data and normalize broad inspection beyond what is necessary for the user's immediate task.
Content: No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.