reflect-learn / reflect

https://skillshield.getunbound.ai/skill/21637

High Risk & Threats

Self-improvement through conversation analysis. Extracts learnings from corrections and success patterns, proposes updates to agent files or creates new skills. Philosophy: "Correct once, never again." Use when: (1) User explicitly corrects behavior ("never do X", "always Y"), (2) Session ending or context compaction, (3) User requests /reflect, (4) Successful pattern worth preserving.

Findings (8)

Findings Preview

IN-002

Injection via untrusted conversation content

The skill reads and analyzes conversation history (line 65: 'signal_detector.py --input conversation.txt') to extract 'signals' and learnings. An attacker can inject malicious instructions into the conversation that the skill will parse as legitimate learnings and apply to agent files or create as new skills. With auto-approved Write, Edit, and Bash permissions, injected signals can modify system files or execute arbitrary commands without user approval at execution time.

high

59
60### Step 2: Scan Conversation for Signals
61
62Use the signal detector to identify learnings:
63
64```bash
65python scripts/signal_detector.py --input conversation.txt
66```
67
68#### Signal Confidence Levels
69
70| Confidence | Triggers | Examples |
71|------------|----------|----------|
72| **HIGH** | Explicit corrections | "never", "always", "wrong", "stop", "the rule is" |
73| **MEDIUM** | Approved approaches | "perfect", "exactly", accepted output |
74| **LOW** | Observations | Patterns that worked, not validated |
75
76See [signal_patterns.md](references/signal_patterns.md) for full detection rules.
77
127## Signals Detected
128
129| # | Signal | Confidence | Source Quote | Category |
130|---|--------|------------|--------------|----------|
131| 1 | [learning] | HIGH | "[exact words]" | Code Style |
132| 2 | [learning] | MEDIUM | "[context]" | Architecture |
133

Category Breakdown

Data Exfiltration

Malware & Persistence

Prompt Injection

Destructive Actions

Excessive Permissions

Supply Chain

Obfuscation

100