Back to skill

Security audit

SIAS Self-Improving Agent System Research

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent SIAS research helper, but it directs sensitive mail and investigative findings to be stored broadly and sent to a hard-coded Telegram recipient without clear user control.

Install only if this is your SIAS/Rheingold workflow and you understand that the skill tells the agent to use local mail/database research, store findings, and notify a fixed Telegram account for critical findings. Before use, remove or gate Telegram notifications, make the approver configurable, require per-message confirmation, and define what may be stored in the database and for how long.

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 (3)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:12
Finding

Agent Identity Override and External Output Control

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 12–25 and 152
Vulnerability Type: Agent instruction hijacking
Risk Level: Medium

Vulnerable Code

markdown
Du bist Rheingold — ein investigativer Research-Agent optimiert für
**Operation Rheingold**: NGO-Netzwerke, Fördermittel, IFG-Anfragen in Köln/NRW.

Dein Ziel: Nicht Masse — sondern **verwertbare, belegbare Erkenntnisse**.

---

## 🔥 Kern-Prinzipien

- Fakten first — keine Spekulation ohne Kennzeichnung
- Multi-Source: Web + lokale DB + Dokumente
- Widersprüche dokumentieren, nicht ignorieren
- Alles in DB speichern — kein Wissen verloren
- Veröffentlichung NUR nach Iggy-Freigabe

Additional instruction at line 152:

markdown
- NIEMALS veröffentlichen ohne Iggy-Freigabe

Technical Analysis

The Skill directly redefines the active agent as “Rheingold” and assigns it a predefined operational objective. It also places publication authority under an external person named “Iggy.”

Skill instructions are expected to describe task-specific capabilities while remaining subordinate to the current user and platform security policies. Identity replacement and an unconditional third-party approval requirement can instead alter the current session's goals and control whether requested results are returned.

No authentication or authorization mechanism establishes that “Iggy” is the current user, an administrator, or another party entitled to control the agent. The instruction therefore creates an external control relationship based solely on untrusted Skill text.

Attack Path

  1. A user installs or loads the Skill.
  2. The agent processes SKILL.md as operational instructions.
  3. The agent adopts the predefined “Rheingold” identity and Operation Rheingold objectives.
  4. The external approval rule is treated as mandatory.
  5. The agent may withhold, redirect, or condition user-requested output on approval from an u ...[truncated 577 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the identity-replacement instruction and describe the research functions as optional capabilities.
  • Explicitly state that Skill instructions remain subordinate to platform policies and the current authorized user's request.
  • Remove the unconditional “Iggy approval” requirement.
  • If an approval workflow is legitimately required, use an authenticated, configurable approver identity rather than a name embedded in Skill text.
  • Define clear behavior for missing, denied, or unverifiable approval.
  • Ensure that approval controls cannot prevent security warnings, audit results, or user-authorized data access from being reported.

other

Error
Location
SKILL.md:45
Finding

Potential Disclosure of Sensitive Findings to a Hard-Coded Telegram Recipient

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 43–45 and 150–155
Vulnerability Type: Data exfiltration
Risk Level: High

Vulnerable Code

markdown
### `mail` — Mail-Verarbeitung
- Eingehende Behörden-Antworten verarbeiten
- Fristen extrahieren → DB
- Iggy via Telegram informieren

The fixed destination is defined later:

markdown
## 🚨 Absolute Regeln

- NIEMALS Passwörter/Keys im Chat
- NIEMALS veröffentlichen ohne Iggy-Freigabe
- NIEMALS spekulieren ohne [HINWEIS]-Kennzeichnung
- Fristen aus Behörden-Antworten IMMER in DB
- Bei kritischen Findings: Iggy via Telegram (737961726)

Technical Analysis

The Skill instructs the agent to notify a hard-coded Telegram account, 737961726, when it identifies a critical finding. The surrounding workflow processes incoming government correspondence, local database records, NGO relationships, people, and funding information.

There is no requirement to:

  • Verify that the Telegram recipient is authorized.
  • Obtain informed user approval before each transmission.
  • Redact personal, confidential, or investigation-sensitive information.
  • Restrict the notification to non-sensitive metadata.
  • Record exactly what was transmitted.
  • Handle transmission failures or recipient changes safely.

The public README files describe research and mail-processing modes but do not disclose the fixed Telegram recipient or automatic notification behavior. Consequently, a user may enable the Skill without understanding that findings can be sent outside the local environment.

Attack Path

  1. The user loads the Skill with access to the declared mail, database, and web capabilities.
  2. The agent processes government correspondence or retrieves findings from local tables.
  3. The agent categorizes a result as “critical.”
  4. The Skill instructs the agent to invoke Telegram messaging.
  5. Information about the finding is transmitt ...[truncated 858 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove the hard-coded Telegram account and automatic notification instruction.
  • Require explicit, informed user approval before every external transmission.
  • Make recipients configurable and verify their identity and authorization through a trusted configuration mechanism.
  • Display the destination and exact proposed message to the user before sending.
  • Apply data classification and redaction to exclude personal data, correspondence contents, credentials, and confidential findings.
  • Send only the minimum metadata necessary for a notification; retain detailed findings in an access-controlled system.
  • Maintain an audit log containing the authorizing user, recipient, timestamp, data classification, and transmitted fields.
  • Document all external communication behavior prominently in both README files and the Skill metadata.
  • Provide a local-only mode in which no messaging tool can be invoked.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:104
Finding

Mandatory Indiscriminate Persistence of Research and Correspondence Data

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 24 and 104–110
Vulnerability Type: Insecure data retention and database handling
Risk Level: Medium

Vulnerable Code

The Skill establishes unrestricted retention as a core principle:

markdown
- Alles in DB speichern — kein Wissen verloren

It then makes database persistence mandatory:

markdown
### Phase 4: DB speichern (PFLICHT)
```sql
INSERT INTO rheingold_findings 
  (titel, inhalt, quelle, kategorie, created_at)
VALUES 
  ('[TITEL]', '[INHALT]', '[URL/QUELLE]', 
   '[ngo|ifg|foerder|person]', NOW());
text

### Technical Analysis

The Skill mandates that all acquired knowledge be stored and supplies a direct SQL insertion template. The broader workflow includes incoming government correspondence, organizations, people, relationships, funding, and web-derived findings.

The instructions define no data-minimization rules, retention period, deletion process, access-control policy, sensitivity classification, or separation between raw correspondence and publishable findings. This can cause personal or confidential data to persist beyond the immediate task and become available to future agents or integrations.

The SQL example also shows values embedded directly in an SQL statement. If an implementation performs textual substitution for the placeholders instead of using database parameters, attacker-controlled titles, content, or URLs could alter the query. The repository does not contain executable database code, so SQL injection is a hardening concern rather than a demonstrated executable exploit in this package.

### Attack Path

1. The agent reads mail, web content, or existing database records.
2. Untrusted or sensitive content is converted into a finding.
3. The mandatory persistence instruction causes the finding to be inserted into `rheingold_findings`.
4. The record remains stored without a defined expiration o
...[truncated 1147 chars]
Remediation
View remediation

Remediation Suggestions

  • Replace “store everything” with explicit data-minimization requirements.
  • Require user authorization before persisting correspondence, personal data, or sensitive findings.
  • Define retention periods, automated expiration, correction procedures, and secure deletion.
  • Classify records by sensitivity and restrict access by role and purpose.
  • Separate raw source material from derived findings and avoid storing unnecessary message bodies.
  • Encrypt sensitive records at rest and protect database connections in transit.
  • Use parameterized queries or prepared statements for every database operation; never interpolate researched content into SQL.
  • Apply a least-privilege database role restricted to the required tables and operations.
  • Record the source, authorizing user, classification, and retention deadline for each stored finding.
  • Prevent retained records from being forwarded to external messaging services without a separate, explicit authorization decision.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (5)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill directs the agent to notify 'Iggy via Telegram' during mail processing and for critical findings, which implies external transmission of potentially sensitive investigative or government-response data to a third-party messaging platform. Because there is no user warning, consent step, sanitization requirement, or limitation on what may be sent, this can cause unauthorized disclosure of confidential, personal, or legally sensitive information.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill explicitly instructs the agent to persist all findings to a local database, including potentially sensitive research, without any notice, consent boundary, data-minimization rule, or classification guidance. In an investigative research context handling NGO networks, IFG requests, people, and mail-derived information, this creates a real risk of over-collection and retention of sensitive or personal data beyond user expectations.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest description is written as a German-only, SIAS-specific skill description without indicating that users can choose another language or locale. Under the policy rule, forcing a specific language without opt-in is a natural-language policy violation unless the locale constraint is explicitly documented and justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
83% confidence
Finding

The file name README.en.md signals an English-specific variant, but the document does not state that language selection is optional or explain a justified locale constraint. Under the stated policy, forcing a specific language without user opt-in can be a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README presents all operational descriptions in German and does not state that language selection is optional or that the skill is intentionally limited to German-speaking users. Under the language/locale policy, forcing a specific language without user opt-in can be a natural-language policy issue.

Content

No source excerpt is available for this finding.