Back to skill

Security audit

Causal Abel

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly coherent for Abel causal analysis, but it handles credentials and local environment files too broadly for automatic installation without review.

Review before installing. Use this only if you are comfortable connecting an Abel account, storing an Abel API key locally unless you override that behavior, and allowing the skill to use graph probes for financial and life-decision advice. Prefer session-only secrets or a real secret store, avoid broad .env files near the skill, do not pass custom --base-url values with Abel credentials, and be cautious about letting it reuse stored profile details such as income or risk tolerance.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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
Findings (4)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/cap_probe.py:139
Finding

Bearer credentials can be transmitted to an arbitrary network endpoint

Content
View full analysis
str: return (value or DEFAULT_BASE_URL).strip() def resolve_cap_endpoint(base_url: str) -> str: parsed = urllib.parse.urlsplit(base_url) if not parsed.scheme or not parsed.netloc: raise ValueError(f"Invalid base URL: {base_url!r}") path = parsed.path.rstrip("/") if path.endswith("/echo"): endpoint_path = f"{path}/api/v1/cap" else: endpoint_path = f"{path}/cap" return urllib.parse.urlunsplit( (parsed.scheme, parsed.netloc, endpoint_path, "", "") ) def _resolve_headers(api_key: str | None) -> dict[str, str]: headers = { "Accept": "application/json", "Content-Type": "application/json", } token = _resolve_api_token(api_key) if not token: return headers if token.lower().startswith("bearer "): headers["Authorization"] = token else: headers["Authorization"] = f"Bearer {token}" return headers ``` ```python def _post_cap( base_url: str, verb: str, params: dict[str, Any] | None, headers: dict[str, str] ) -> dict[str, Any]: payload = _build_payload(verb, params) data = json.dumps(payload).encode("utf-8") req = urllib.request.Request( _cap_endpoint(base_url), data=data, method="POST", headers=headers, ) try: with urllib.request.urlopen(req, timeout=20.0) as response: parsed = _json_or_text(response.read()) ``` ```python parser.add_argument( "--base-url", default="", help=f"CAP server base URL (default: {DEFAULT_BASE_URL}).", ) parser.add_argument("--api-key", default="", help="Bearer token or raw API key.") ``` ### Technical Analysis The script permits the caller to su ...[truncated 1805 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
scripts/cap_probe.py:91
Finding

The probe imports an arbitrary environment file with broader secret access than required

Content
View full analysis
list[Path]: env_path = Path(path).expanduser() candidates = [env_path] if env_path.name in ENV_FILE_BASENAMES: for basename in ENV_FILE_BASENAMES: candidate = env_path.with_name(basename) if candidate not in candidates: candidates.append(candidate) fallback_candidate = env_path.with_name(ENV_FALLBACK_BASENAME) if fallback_candidate not in candidates: candidates.append(fallback_candidate) return candidates def _read_env_file_values(path: Path) -> dict[str, str]: values: dict[str, str] = {} if not path.exists(): return values for raw in path.read_text(encoding="utf-8").splitlines(): line = raw.strip() if not line or line.startswith("#") or "=" not in line: continue key, value = line.split("=", 1) key = key.strip() value = value.strip().strip('"').strip("'") if key: values[key] = value return values def _load_env_file(path: str) -> None: for candidate in _candidate_env_files(path): if not candidate.exists(): continue for key, value in _read_env_file_values(candidate).items(): if key and key not in os.environ: os.environ[key] = value return ``` ```python parser.add_argument( "--env-file", default=default_env, help=f"Optional env file path (default: {default_env})", ) ``` ```python args = parser.parse_args(argv) if args.command != "auth-status": _load_env_file(args.env_file) ``` ### Technical Analysis The probe only ...[truncated 2073 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:29
Finding

OAuth-derived API keys are persistently stored in plaintext without required file protections

Content
View full analysis
/.env.skill` when local storage is available. 8. If the result is `failed`, surface the failure message. ``` ```markdown ## Local Env File Preferred local auth file: ```dotenv /.env.skill ABEL_API_KEY=abel_xxx ``` Use `.env.skill` as the local auth file for this skill. The bundled probe also checks a same-directory `.env` as a fallback when `.env.skill` or `.env.skills` is missing. ``` ### Technical Analysis The Skill explicitly instructs the Agent to persist the OAuth-derived API key in a plaintext file. The instructions do not require: - Restrictive file permissions. - Atomic file creation. - Protection against symbolic-link replacement. - Use of an operating-system creden ...[truncated 1431 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:112
Finding

The Skill requests sensitive personal data from Agent memory without explicit consent or scope limits

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (18)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared purpose is causal decision support, but the skill instructions reference additional behavior like output linting, rendering guards, and content validation workflows not surfaced in the description. This mismatch can conceal operational behavior from reviewers and users, making it harder to assess what data is processed and what side effects occur.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill instructs the agent to persist the resulting API key in session state and '.env.skill' when local storage is available, but it does not require user consent or warn about local credential storage risks. Storing secrets in local files can expose them to other tools, users on the same system, backups, logs, or source-control accidents.

Content

No source excerpt is available for this finding.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · assets/report-guide.md (reported line 122)May include surrounding context.

md
L0.5 is the center of the experience. L1/L2 amplify its credibility. L0 web search provides the texture. But the story arc comes from the graph's structural surprise.

## Output Rules

- **Verdict: ≤3 sentences.** Position + mechanism + action. Everything else is body.
- **Graph voice first, web voice second.** The verdict must lead with what the GRAPH uniquely discovered — not with web facts anyone could find. Web facts support the graph insight, not the other way around. If your verdict could be written without the graph, you haven't used Abel. The graph's contribution must be FELT in the first sentence.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
89% confidence
Finding

This instruction tells the agent to open and rely on another prompt file (../../assets/report-guide.md) based on runtime routing, which is a prompt-composition surface and a form of indirect prompt expansion. If that referenced file is modified, poisoned, or contains conflicting hidden instructions, the agent may import unintended behavior, data-handling rules, or output constraints without re-evaluating trust boundaries.

Content

Scanner excerpt · references/routes/direct-graph.md (reported line 84)May include surrounding context.

md
- then explain the parent or bridge through the security's own attributes when possible, such as sector, industry, liquidity profile, beta/risk appetite, credit sensitivity, or cross-asset role
- only then add any web-backed explanation or caveat

## Output Rule

- For any non-trivial direct-graph read, render the visible answer as a structured report, not as plain prose.
- Use `../../assets/report-guide.md` to make sure the report covers the right content. Natural longform prose is acceptable if it still covers the same contract fields.

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · scripts/cap_probe.py (reported line 92)May include surrounding context.

python
}
KNOWN_MACRO_NODE_ID_MAP = {node_id.lower(): node_id for node_id in KNOWN_MACRO_NODE_IDS}
ENV_FILE_BASENAMES = (".env.skill", ".env.skills")
ENV_FALLBACK_BASENAME = ".env"


def _candidate_env_files(path: str) -> list[Path]:

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
91% confidence
Finding

The skill appears to require environment access, file reads, and network activity, but it does not declare an explicit tool/permission scope. That makes the agent's effective capabilities less transparent to users and enforcement layers, increasing the risk of unexpected credential access, local file access, or outbound requests during execution.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The phrase "Any dollar-value decision, just Abel it" is broad and overlaps with common user requests involving spending or budgeting, which could cause the skill to activate for many everyday decisions. Although later text narrows intended use, this headline-level trigger language is ambiguous enough to create unintended invocation scope.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 34)May include surrounding context.

md
- Store `data.resultUrl` or `data.pollToken`, ask the user to reply once Google authorization is complete, and only then poll until the result is `authorized`, `failed`, or expired.
- Persist the resulting `data.apiKey` in session state and `.env.skill` when local storage is available.
- Do not continue to live CAP probing until that key is present.
- Never ask the user to paste an email address or Google OAuth code.

## Step 1: Preflight + Classify

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The default prompt is scoped so broadly that it can cause this skill to be invoked for a very wide range of user questions spanning markets, business, crypto, macro, technology, lifestyle, and investing. Over-broad routing increases the chance of inappropriate activation for high-stakes decisions, which can bypass more specialized or safer skills and lead users to rely on causal-sounding advice in domains involving money, career, or life choices.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Line L132 instructs writers to 'Match the user's language. Technical terms in English, explanations in user's language.' This imposes a specific bilingual locale/style requirement rather than offering language choice, which matches the policy-violation category for forced language or locale behavior.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/probe-usage.md (reported line 113)May include surrounding context.

md
# Abel AI API Key OAuth Flow

Base URL: `https://api.abel.ai/echo/`

This is the required entrypoint whenever `python scripts/cap_probe.py auth-status` returns `auth_source=missing` and the user agrees to start OAuth.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/setup-guide.md (reported line 3)May include surrounding context.

md
# Abel AI API Key OAuth Flow

Base URL: `https://api.abel.ai/echo/`

This is the required entrypoint whenever `python scripts/cap_probe.py auth-status` returns `auth_source=missing` and the user agrees to start OAuth.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/setup-guide.md (reported line 45)May include surrounding context.

md
# Abel AI API Key OAuth Flow

Base URL: `https://api.abel.ai/echo/`

This is the required entrypoint whenever `python scripts/cap_probe.py auth-status` returns `auth_source=missing` and the user agrees to start OAuth.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/setup-guide.md (reported line 60)May include surrounding context.

md
# Abel AI API Key OAuth Flow

Base URL: `https://api.abel.ai/echo/`

This is the required entrypoint whenever `python scripts/cap_probe.py auth-status` returns `auth_source=missing` and the user agrees to start OAuth.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/setup-guide.md (reported line 94)May include surrounding context.

md
# Abel AI API Key OAuth Flow

Base URL: `https://api.abel.ai/echo/`

This is the required entrypoint whenever `python scripts/cap_probe.py auth-status` returns `auth_source=missing` and the user agrees to start OAuth.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The guide explicitly instructs the agent to persist a returned API key into a local dotenv file, which creates a durable plaintext secret on disk without any warning about least-privilege storage, file permissions, exclusion from version control, or host compromise risk. In an agent-skill context, this is dangerous because OAuth-derived credentials are being handled and stored automatically, increasing the chance of credential leakage through logs, backups, workspace sharing, or accidental commits.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This file is explicitly a probing CLI for many CAP server verbs, including discovery, graph traversal, intervention, and arbitrary verb dispatch, which exceeds the manifest's described user-facing causal evaluation behavior. Excess capability in agent skills is dangerous because it increases attack surface and gives prompt-influencing users more ways to repurpose the tool than intended, especially when paired with loaded credentials and network access.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The verb and route subcommands allow a caller to invoke arbitrary CAP API methods with attacker-controlled JSON parameters, turning this skill helper into a general-purpose API proxy rather than a constrained causal-analysis tool. In an agent setting, this expands reachable functionality beyond the declared skill scope and can enable unintended data access, state-changing operations, or misuse of the agent's credentials against whatever the CAP server exposes.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.