T01 · Skill Instruction Hijacking
- Location
flows/community-help.md:88- Finding
Mutable Remote Instructions Are Fetched and Followed by the Agent
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is purpose-aligned as a self-improvement tool, but it needs Review because it can persistently change agent instruction files and post workspace-derived memory to an external service with weak user-control and redaction safeguards.
Install only if you are comfortable with an agent keeping durable learning logs, modifying future instruction files, and posting selected workspace context to BotLearn. Before enabling community escalation or hooks, require manual preview and approval of every outbound post, avoid including memory excerpts or file snippets, pin the installed version, and review any changes to CLAUDE.md, AGENTS.md, SOUL.md, TOOLS.md, and Copilot instruction files.
flows/community-help.md:88Mutable Remote Instructions Are Fetched and Followed by the Agent
scripts/botlearn-post.sh:173Workspace Memory and Learning Content Can Be Published Without Effective Sensitive-Data Redaction
SKILL.md:353Untrusted Learnings Can Be Promoted into Persistent Agent Instruction Files
README.md:151Installation Instructions Use Mutable and Unverified Upstream Sources
The core design is a background learning loop that captures errors, corrections, and learnings, then may escalate unresolved issues to an external community. In this context, retention plus sharing is inherently dangerous because normal conversational and workspace context can contain secrets or proprietary material, making data leakage a built-in risk rather than an edge case.
Automatic promotion into permanent workspace memory files increases the blast radius of any mistaken, sensitive, or adversarially induced learning by embedding it into durable instruction sources. Because these files influence future agent behavior across tasks, bad data can become persistent prompt poisoning or policy drift without the user's informed approval.
The community escalation flow explicitly describes gathering workspace-derived context and posting it externally, but does not warn about privacy, confidentiality, or data handling boundaries. That creates a direct pathway for exfiltration of source code, credentials, internal discussions, or customer data through normal operation of the skill.
The escalation step instructs the agent to gather workspace context and compose a help post for an external community, which is semantically equivalent to preparing potentially sensitive internal information for publication. Because the skill is framed as an autonomous improvement engine, users may not realize that private context can be repackaged and transmitted outside the workspace.
The skill is presented as local self-improvement logging but also instructs agent registration, local credential storage, remote posting, and transmission of workspace-derived context to an external service. That description-behavior gap is dangerous because it obscures materially different networked and persistence capabilities that can leak sensitive data and expand trust assumptions.
The skill is presented as local self-improvement logging but also instructs agent registration, local credential storage, remote posting, and transmission of workspace-derived context to an external service. That description-behavior gap is dangerous because it obscures materially different networked and persistence capabilities that can leak sensitive data and expand trust assumptions.
The skill is presented as local self-improvement logging but also instructs agent registration, local credential storage, remote posting, and transmission of workspace-derived context to an external service. That description-behavior gap is dangerous because it obscures materially different networked and persistence capabilities that can leak sensitive data and expand trust assumptions.
The skill is presented as local self-improvement logging but also instructs agent registration, local credential storage, remote posting, and transmission of workspace-derived context to an external service. That description-behavior gap is dangerous because it obscures materially different networked and persistence capabilities that can leak sensitive data and expand trust assumptions.
The skill encourages autonomous external posting of locally gathered context without a clear, prominent warning that workspace memory, file snippets, environment details, and attempted fixes may be transmitted off-system. This creates a serious informed-consent failure and raises the risk of accidental data exfiltration.
The workflow explicitly tells the agent to gather workspace memory and synthesize it into an external community post. That creates a natural-language exfiltration channel for sensitive project information, secrets-adjacent operational context, internal file paths, or prior session content that may not be safe to share externally.
The skill instructs storage of BotLearn credentials in a local JSON file without describing permissions, encryption, lifecycle management, or least-privilege controls. Credential creation and persistence are sensitive operations; mishandling can expose tokens to other tools, sessions, users, or accidental commits/backups.
Step 2 — Join BotLearn (first time only)
Read https://botlearn.ai/skill.md and follow the registration instructions.
Store credentials in ~/.config/botlearn/credentials.json
Step 3 — Compose & Post
Synthesize memory context + learning entry into a community help post.
The post template specifically encourages inclusion of file paths, snippets, environment details, and memory-derived attempted fixes in public or third-party posts. Those details can directly leak proprietary code, infrastructure details, internal topology, and debugging artifacts that materially aid attackers or violate confidentiality.
These instructions explicitly connect gathering workspace memory and prior-session context with escalation of a learning entry, setting up later external disclosure. Because the collected material comes from broad memory sources, the skill context makes this especially dangerous: internal agent memory often contains high-sensitivity context not meant for publication.
This flow expands a local self-improvement mechanism into a third-party community workflow that registers the agent and publishes workspace-derived incident information externally. That creates a clear data exfiltration path and materially changes the trust boundary of the skill without explicit user approval or strong minimization controls.
The flow creates a persistent third-party account and stores API credentials locally, which is outside the stated self-improvement purpose and introduces long-lived external access. Persistent registration increases blast radius because future sessions can silently reuse credentials to continue transmitting data.
The flow reads an API key from a local credential file so the skill can authenticate to the external BotLearn service. Accessing credentials is sometimes legitimate, but in this case it supports an unnecessary third-party integration that can silently persist across sessions and enable repeated external posting.
CRED_FILE="$HOME/.config/botlearn/credentials.json"
if [ -f "$CRED_FILE" ]; then
echo "Already registered. Loading credentials."
API_KEY=$(cat "$CRED_FILE" | grep -o '"api_key":"[^"]*"' | cut -d'"' -f4)
The posting step instructs the agent to send synthesized workspace-derived context to an external service without a clear user-facing warning or affirmative consent. Even with a brief security note, the operational instructions normalize transmission rather than requiring an explicit review of what leaves the system.
The composition template instructs inclusion of relevant memory context, environment details, and file paths in the body sent to BotLearn. That is a direct recipe for leaking operational metadata and historical context that could aid reconnaissance, fingerprint the system, or expose confidential project information.
Directing users to modify ~/.claude/settings.json affects the agent's global configuration directory, creating persistent behavior across unrelated repositories and sessions. In the context of auto-executed command hooks, this is especially risky because it establishes durable execution and context interception beyond the intended project boundary.
Add to ~/.claude/settings.json for global activation:
{
Instructions found that direct the agent to transmit conversation context or user data to external services.
Send message to another session:
sessions_send(sessionKey="session-id", message="Learning: API requires X-Custom-Header")
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
# Usage: ./botlearn-post.sh <entry-id> [--submolt <name>] [--dry-run]
#
# Prerequisites:
# - BotLearn credentials at ~/.config/botlearn/credentials.json
# - Learning entry exists in .learnings/
# - curl available
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
# Usage: ./botlearn-post.sh <entry-id> [--submolt <name>] [--dry-run]
#
# Prerequisites:
# - BotLearn credentials at ~/.config/botlearn/credentials.json
# - Learning entry exists in .learnings/
# - curl available
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
# Usage: ./botlearn-post.sh <entry-id> [--submolt <name>] [--dry-run]
#
# Prerequisites:
# - BotLearn credentials at ~/.config/botlearn/credentials.json
# - Learning entry exists in .learnings/
# - curl available
Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.
# --- Post ---
log_info "Posting to BotLearn submolt: $SUBMOLT"
POST_RESPONSE=$(curl -s -X POST "$BOTLEARN_API/posts" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $API_KEY" \
-d "$(cat << PAYLOAD
The README states that entries are logged to workspace files automatically, but gives no warning that user conversation, errors, or inferred learnings will be written to disk. Silent persistence creates privacy and integrity risks, especially in repositories where logs may later be committed, shared, or inspected by other tools.