AAK-TOXICFLOW-001
Toxic flow: sensitive source paired with external sink
| Field | Value |
|---|---|
| Severity | HIGH |
| Category | TOOL_POISONING |
| Shipped | v0.3.5 (2026-04-25) — feature-flagged |
| Scanner | agent_audit_kit/scanners/toxic_flow.py |
| Data | agent_audit_kit/data/toxic_flow_pairs.yml |
| OWASP MCP | none |
| OWASP Agentic | ASI02 (Tool Misuse and Exploitation) |
| AICM | AIS-12, CCC-08 |
| Origin | Snyk Agent Scan parity |
Feature-flag status
This rule is off by default. Set AAK_TOXIC_FLOW=1 in the
environment to opt in. It pairs sources and sinks from a static list,
so it over-reports on a project that legitimately does both; the
deny-graph design review that would replace the list has not happened.
What it catches
An agent project exposes both a sensitive source tool (filesystem
read, secrets read, database query) and an external sink tool
(HTTP POST, email send, git push, shell exec). Even if each tool is
individually safe, the LLM can chain them — the canonical exfiltration
pattern is read_file -> http.post.
The scanner builds a per-scan tool graph from:
- MCP server names + arg strings in
.mcp.json/.cursor/mcp.json/.vscode/mcp.json/.amazonq/mcp.json/mcp.json - Function names of
@tool/@mcp.tool/@server.tooldecorated Python functions in the repo, bare or with empty parentheses
A pair is reported against the config that declared its first configured
tool or, when every tool in it is a repo function, at the def of its
first tool.
Tool identifiers are matched against agent_audit_kit/data/toxic_flow_pairs.yml
which lists known source families (fs_read, secrets_read,
db_query) and sink families (http_post, email_send, git_push,
shell_exec) plus the pairs that fire.
Suppression
Add the source/sink pair to .aak-toxic-flow-trust.yml with a
non-empty justification::
trust:
- source: fs_read
sink: http_post
justification: "Documented data-export feature; sink is allow-listed in egress proxy."
A trust entry without a justification does not suppress (we want the audit trail to carry the reason, not just the decision).
What it does NOT catch
- Implicit toxic flows where the LLM bridges two non-tool capabilities (e.g. system-prompt context + a shell pipe). Prompt-level controls belong elsewhere.
- Pairs not yet enumerated in
toxic_flow_pairs.yml. Submit a PR with the new family + cite the canonical incident. - Tools declared in other languages (TS / Java / Rust): only Python decorators are read.
Remediation
Three knobs, in priority order:
- Remove the pairing. Drop the source or the sink if the agent does not actually need both at once.
- Scope the source. Restrict
fs_read/db_queryto a directory or schema the sink cannot reach. - Allow-list with justification. Add the pair to
.aak-toxic-flow-trust.ymland document the egress controls that make the chain safe.
Registry reference
This section is generated from the rule registry (agent_audit_kit/rules/builtin.py) when the documentation is built, so it follows the scanner. The rest of the page is hand-written; where the two disagree, the registry is what runs.
- Severity
- HIGH
- Category
- Tool Poisoning (
tool-poisoning) - Scanner
agent_audit_kit/scanners/toxic_flow.py- Auto-fixable
- No.
- SARIF name
ToxicFlowSourceSink
What it detects
An agent project exposes both a sensitive source tool (filesystem read, secrets read, database query) and an external sink tool (HTTP POST, email send, git push) without an explicit .aak-toxic-flow-trust.yml allow-list entry. Even if each tool is individually safe, the LLM can chain them — the canonical exfil pattern is read_file -> http.post. Suppress with an allow-list when the pairing is a documented product feature.
How to fix
Add the source/sink pair to .aak-toxic-flow-trust.yml with a justification: field, scope the source tool to a directory the sink cannot reach, or remove one side of the pair. The finding's evidence names the families and the tools on each side.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-TOXICFLOW-001
Limitations
Off unless AAK_TOXIC_FLOW=1 is in the environment. It reads the server names and args strings of .mcp.json, .cursor/mcp.json, .vscode/mcp.json, .amazonq/mcp.json and mcp.json at the scan root, and the names of Python functions decorated @tool, @mcp.tool or @server.tool with no arguments, and matches each against the names in toxic_flow_pairs.yml. A server is judged by its name and arguments, not by the tools it lists at run time, so @modelcontextprotocol/server-filesystem, whose read tool is read_text_file, is no source. A decorator with arguments (@tool("read_file"), @mcp.tool(name=...)), a TypeScript server.tool(...) and a skill under .claude/skills/ declare no tool. A Python file over 512 KB, or an MCP config or trust file over 1 MB, is not read.
References
- OWASP Agentic Top 10: ASI02 (Tool Misuse and Exploitation)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-CHAIN-01
- CSA AI Controls Matrix (AICM): AIS-12, CCC-08