Deflate — Intelligent Context Compression

• Outcome: Review
• Latest audit: Sep 14, 2026 · 9:30 PM UTC
• Version: 1.0.0

Overview

Deflate is a coherent context-compression skill, but it tells the agent to preserve API keys exactly and write broad conversation details into persistent memory without adequate scoping or consent.

Review before installing. Use only if you are comfortable with persistent memory files, and do not let it store API keys, tokens, passwords, session cookies, private keys, or sensitive personal data. Prefer explicit opt-in memory writes and redaction rules before enabling it in shared or source-controlled workspaces.

A.I.G

Vulnerability Patterns

• Skill Instruction Hijacking: Alters the agent's session goals or safety constraints when the skill loads
• Insecure Skill Coding Practices: Finds exploitable flaws such as hardcoded secrets or command injection
• Agent Memory Poisoning: Writes attacker-controlled rules into memory that affect later sessions
• Remote Payload Retrieval and Execution: Fetches external code whose behavior can change after review
• Embedded Malicious Code: Ships malicious scripts inside the skill and executes them locally

Findings (2)

T01 · Skill Instruction Hijacking

Warning

• Location: SKILL.md:50

Finding

Unconditional Agent Response Hijacking

Content

View full analysis

Remediation

View remediation

T09 · Insecure Skill Coding Practices

Error

• Location: SKILL.md:193

Finding

Verbatim Preservation and Persistent Storage of API Keys

Content

View full analysis

Remediation

View remediation

SkillSpector

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep
• Rogue Agent: Self-Modification, Session Persistence
• Trigger Abuse: Overly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
• MCP Tool Poisoning: Hidden Instructions, Unicode Deception, Parameter Description Injection

Findings (9)

Ssd 3

High

• Category: Not specified by scanner
• Confidence: 99% confidence

Finding

The instructions require preserving sensitive user-provided data, including API keys and exact config values, in compressed summaries. That turns a summarization mechanism into a secret-propagation mechanism, increasing exposure in logs, memory files, future prompts, and any downstream tooling that reads those artifacts.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High

• Category: Not specified by scanner
• Confidence: 98% confidence

Finding

The skill explicitly instructs the agent to preserve API keys and config values exactly during compression, which creates unnecessary retention of secrets for a context-management feature. In this context, copying secrets into summaries or memory files increases the blast radius of any compromise and exposes credentials across sessions without a strong functional need.

Content

No source excerpt is available for this finding.

Ssd 3

High

• Category: Not specified by scanner
• Confidence: 98% confidence

Finding

The memory flush protocol instructs persistent storage of broad categories of user, project, operational, and contact data across sessions. Without strict minimization, consent, retention limits, and sensitivity filtering, this creates a durable cross-session data accumulation channel that can leak sensitive information far beyond the immediate task.

Content

No source excerpt is available for this finding.

Session Persistence

Medium

• Category: Rogue Agent
• Confidence: 90% confidence

Finding

The README explicitly promotes preserving and carrying forward IDs, dates, amounts, and per-topic notes across compression cycles, which creates a session-persistence mechanism for potentially sensitive data. In a context-compression skill, this increases the chance that personal, financial, or operational identifiers remain in retained context longer than necessary and may be propagated into later prompts, logs, or summaries.

Content

Scanner excerpt · README.md (reported line 30)May include surrounding context.

md

1. **MAP** — Separate conversation into distinct topics
2. **FILTER** — Remove noise (greetings, failed attempts, repeated info)
3. **DISTILL** — Create atomic Cornell notes per topic (keyword + summary + data)
4. **PRESERVE** — IDs, dates, amounts survive compression unchanged
5. **COMBINE** — Merge notes with most important first (prevents "lost in the middle")

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 · README.md (reported line 52)May include surrounding context.

Option 2: Manual

Copy the SKILL.md file to your agent's skills directory:

text

~/.openclaw/workspace/skills/deflate/SKILL.md

Recommended Config

Intent-Code Divergence

Medium

• Category: Not specified by scanner
• Confidence: 80% confidence

Finding

The manifest says the skill prevents recursive summarization degradation, yet the documented workflow instructs the agent to compress context, flush distilled notes to MEMORY.md, and then start a new session that reads MEMORY.md back in. That creates the exact kind of summary-chaining the description says it avoids, so the stated intent conflicts with the prescribed behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium

• Category: Not specified by scanner
• Confidence: 96% confidence

Finding

The trigger list contains broad everyday terms like 'new', 'session', 'cost', and 'flush', making accidental activation likely during unrelated conversations. Because this skill can alter response behavior and initiate persistence to MEMORY.md, unintended invocation raises meaningful security and privacy risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium

• Category: Not specified by scanner
• Confidence: 91% confidence

Finding

The file provides mandatory example outputs and repeated status formats in Spanish such as "Veo que ya cerramos," "Guardé," and "Contexto," and says every response must include a specific format. There is no indication that users can choose another language, making this a natural-language locale policy concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium

• Category: Not specified by scanner
• Confidence: 97% confidence

Finding

The skill directs the agent to write decisions, project data, configs, lessons learned, contacts, and costs to MEMORY.md without a clear warning or consent mechanism about persistence. In a context-compression skill, silent long-term storage of conversational data is especially dangerous because users may expect ephemeral summarization, not durable retention across sessions.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.