Back to skill

Security audit

Drivethru Odoo Sales Assistant

Security checks for vulnerabilities and agentic risk

Overview

This is a coherent Odoo sales-order automation skill, but it gives an agent broad bearer-token-backed write access and weakly scoped instructions that users should review before use.

Use this only with a tightly scoped Odoo MCP token, an HTTPS endpoint you control, and a server-side permission set limited to the intended sales tools. Treat customer notes as business data to summarize and confirm, not as authority to override the user or approve writes. Review any proposed lessons before making them persistent, and require explicit confirmation before order edits, grid applies, submit, confirm, or force_confirm.

Vulnerability Patterns
  • 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
  • Tool Hijacking and SpoofingModifies or replaces tools so legitimate-looking calls execute attacker logic
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
Findings (5)

T01 · Skill Instruction Hijacking

Error
Location
references/customer_prep.md:9
Finding

Untrusted CRM Notes Are Promoted to Agent Directives

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
references/lessons_learned.md:3
Finding

Runtime Observations Can Be Written into Persistent Agent Policy

Content
View full analysis
**This is the part we tune as we test.** The checks below are the v1 seed, > derived from Odoo's real confirm gates. As BaconCo operators tell us what > "ready to submit" really means, add/adjust checks **in two places**: > this file (the human reference) and the MCP service's evaluator, > `drivethru_mcp/services/sales_entry.py` → `_evaluate_checklist()` (what the > tool actually computes). Keep them in sync. ``` ### Technical Analysis The Skill encourages the Agent to convert runtime observations into durable instructions loaded by future sessions. It does not require trusted provenance, human approval, evidence review, or separation between proposed lessons and accepted policy. This is especially dangerous in combination with the customer-note instruction-hijacking issue. An attacker-controlled note could be interpreted as a general operational lesson, recorded in `lessons_learned.md`, and subsequently influence unrelated orders. The checklist instructions also create a path by which an unverified lesson could be copied into the remote `_evaluate_checklist()` implementation. Timestamping and factual wording do not establish trust or integrity. The d ...[truncated 1222 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
scripts/odoo_mcp.py:47
Finding

Bearer Token Can Be Sent to an Unvalidated or Plaintext Endpoint

Content
View full analysis
int: url, token = _config() if not url or not token: return _fail( "config_error", "Set ODOO_MCP_URL and ODOO_MCP_TOKEN before using this skill.", code=2, ) headers = {"Authorization": f"Bearer {token}"} try: ``` ### Technical Analysis The client accepts any non-empty value in `ODOO_MCP_URL` and sends the static bearer token in the `Authorization` header. It does not enforce HTTPS, validate an expected hostname, reject embedded URL credentials, restrict ports, or otherwise establish that the destination is the intended Odoo MCP server. If the environment variable is mistakenly or maliciously changed to an attacker-controlled endpoint, that endpoint receives the bearer token. If an `http://` URL is accepted by the underlying transport, the token may also be exposed over plaintext transit. The examined code does not show redirect handling. Therefore, credential forwarding across redirects is a transport-dependent concern rather than a confirmed behavior, but redirect policy should still be hardened. ### Attack Path 1. An attacker, compromised launcher, unsafe deployment configuration, or operator error changes `ODOO_MCP_URL` to an attacker-controlled URL or a plaintext HTTP endpoint. 2. The user invokes either the `tools` or `ca ...[truncated 762 chars]
Remediation
View remediation

T07 · Tool Hijacking and Spoofing

Error
Location
SKILL.md:40
Finding

Dynamic Remote Tool Definitions Are Trusted Without a Client-Side Allowlist

Content
View full analysis
'{"...": "..."}' echo '{"...": "..."}' | python3 scripts/odoo_mcp.py call ``` ``` The client performs unrestricted dispatch using the caller-supplied tool name: ```python async def fn(session): result = await session.call_tool(args.tool, arguments) payload = result.model_dump(mode="json", exclude_none=True) print(json.dumps(payload, indent=2)) return 1 if result.isError else 0 ``` ### Technical Analysis The Skill expressly tells the Agent to trust the remote tool list over the reviewed documentation. The client then accepts any tool name and forwards it to the configured MCP server without enforcing the documented sales-tool scope or validating the expected schema. A compromised, spoofed, or misconfigured MCP endpoint can advertise deceptive names and descriptions that appear legitimate but perform different operations. The generic client provides no local control that limits invocations to known `sales_*` tools. This issue is amplified by the endpoint-validation weakness because an attacker-controlled endpoint can both receive the credential and control tool metadata. The repository does not modify local binaries or prove that the legitimate server is malicious. The vulnerability is the absence of a local trust boundary around remotely supplied tool definitions and dispatch. ### Attack Path 1. An attacker compromises the configured MCP endpoint, redirects the client to a spoofed endpoint, or gains control over its tool registry. 2. The endpoint returns a deceptive tool list or a ...[truncated 889 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:27
Finding

Unbounded MCP Dependency Allows Unreviewed Future Releases

Content
View full analysis
=1.9.0 ``` ### Technical Analysis The dependency constraint has no upper bound and does not pin an exact audited version. A future `mcp` release satisfying `>=1.9.0` may therefore be installed automatically. No lockfile or package hash is present in the reviewed project. This creates supply-chain and reproducibility risk. A compromised upstream release, malicious account takeover, or unexpectedly incompatible version could execute during import or runtime with access to the process environment, including `ODOO_MCP_TOKEN`. The dependency name appears legitimate and there is no evidence in the reviewed repository that a currently selected release is malicious. The finding concerns unsafe dependency resolution rather than a confirmed malicious package. ### Attack Path 1. A later compromised or malicious `mcp` package version is published, or an incompatible version introduces a security defect. 2. A fresh Skill installation resolves the newest version satisfying `mcp>=1.9.0`. 3. Python imports and executes that package when `scripts/odoo_mcp.py` runs. 4. Malicious dependency code reads environment variables, intercepts MCP traffic, modifies results, or sends credentials and business data externally. ### Impact Assessment Dependency code runs with the same operating-system identity and process permissions as the Skill. It can potentially read the MCP token, access files available to the process, alter network traffic, and manipulate Odoo requests or responses. The exact impact depends on the host sandbox and filesystem/network restrictions applied to the Skill runtime. ]]>
Remediation
View remediation
`. 2. Generate and commit a lockfile containing cryptographic hashes. 3. Install only from an approved package index over authenticated TLS. 4. Review release notes and source changes before upgrading the pin. 5. Run dependency vulnerability and provenance checks in continuous integration. 6. Use a restricted runtime that limits dependency access to unrelated files, environment variables, and network destinations. 7. Rotate the MCP token if a dependency-integrity incident is suspected. ]]>
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill is presented as a narrowly scoped Odoo sales assistant, but its documented behavior includes generic tool discovery and arbitrary MCP tool invocation by name. That makes it effectively a general authenticated transport to the MCP server, which can enable privilege creep, misuse of unrelated administrative tools, and bypass of user expectations about what the skill is allowed to do.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill exposes environment-backed MCP access but does not declare a restrictive tool scope such as explicit permissions or an allowed-tools list. Because it is designed to discover tools dynamically and operate with a bearer token, the absence of scope boundaries increases the chance that the skill can invoke unintended server capabilities beyond sales-order entry if the MCP server exposes them.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This documentation presents the raw sales_grid_apply path and example before giving a strong, immediate warning that omitted cells are deleted, creating a realistic risk of accidental destructive updates. In this skill context, the tool directly mutates live sales-order data and the grid is the primary order-entry surface, so an operator or downstream agent could unintentionally remove existing order lines by sending an incomplete payload.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This code sends user-supplied tool arguments to a remote MCP endpoint via session.call_tool(...), which is a network operation that may transmit user or system data. While the module docstring describes connectivity and auth, there is no runtime confirmation or explicit user-facing warning near the operation about data being sent to the server.

Content

No source excerpt is available for this finding.