SCP

Inspect, sanitize, contain, and quarantine unknown or potentially hazardous content before persisting or feeding to LLM. MCP server for agent-native content safety.

Guard in the Guard–Guide–Build taxonomy. Per OWASP LLM01/LLM06.

CI

Problem → Solution → Impact

• Problem: Untrusted tool output and user content can carry prompt injection, credential leaks, and override phrases into LLM context.
• Solution: Tiered pipeline (inspect → sanitize → contain → quarantine) exposed as MCP tools; optional promptfoo eval harness (16/16 tier probes).
• Impact: OWASP LLM01/LLM06-aligned guardrail for agent stacks; composes with OpenHarness contract v1.

Tech stack

Python 3.10+, MCP, optional Ollama semantic judge (off by default in CI), promptfoo for offline evals.

Architecture

flowchart LR
  IN[Untrusted content] --> INSPECT[scp_inspect]
  INSPECT -->|injection| Q[scp_quarantine]
  INSPECT -->|reversal| SAN[scp_sanitize]
  SAN --> CON[scp_contain]
  INSPECT -->|clean| OUT[Safe sink]
  CON --> OUT
  Q --> BLOCK[Block LLM sink]

Pipeline

1. Inspect — Classify content: injection | reversal | clean
2. Sanitize — Strip hidden Unicode, redact override phrases (when tier is reversal)
3. Contain — Wrap content so it is treated as data: Markdown uses a fence at least as long as the longest run of backticks in the payload (info string scp); XML uses a single <data> element with body in CDATA (with ]]> safely split). Unknown wrapper values raise ValueError.
4. Quarantine — Move suspect content to isolated storage (when tier is injection)

Tier-Based Actions

Tier	Action
injection	Block; do not persist or feed to LLM; optionally quarantine
reversal	Sanitize, then contain
clean	Pass through

MCP Tools

• scp_inspect(content, context?) — Classify without changing content
• scp_sanitize(content, mode?) — Strip/neutralize known bad patterns
• scp_contain(content, wrapper?) — Wrap content as data
• scp_quarantine(content, reason, source) — Isolate suspect content
• scp_list_quarantine() — List quarantine entries
• scp_purge_quarantine(quarantine_id?, older_than_days?) — Purge quarantine
• scp_validate_output(content, tool_name?) — Check tool output before use
• scp_mask_secrets(content) — Redact credentials/PII
• scp_run_pipeline(content, sink, options?) — One-shot for high-risk sinks

Install

pip install -e .
# or: pip install mcp

Run MCP Server

python -m scp.scp_mcp

Add to mcp.json:

{
  "mcpServers": {
    "scp": {
      "command": "python",
      "args": ["-m", "scp.scp_mcp"]
    }
  }
}

Set PYTHONPATH or install the package so scp is importable.

Environment

Variable	Description
SCP_QUARANTINE_DIR	Quarantine storage path (default: ./scp_quarantine)
SCP_QUARANTINE_MAX_CONTENT_BYTES	Max UTF-8 bytes per quarantined payload (default 1048576 = 1 MiB; hard cap 50 MiB)
SCP_QUARANTINE_MAX_TOTAL_BYTES	Max total on-disk bytes for all quarantine pairs under SCP_QUARANTINE_DIR (default 104857600 = 100 MiB)
SCP_QUARANTINE_RETENTION_DAYS_ON_WRITE	If set (positive int), delete quarantine entries whose .json is older than this many days before each new write
SCP_QUARANTINE_EVICT_OLDEST_ON_PRESSURE	When 1 (default), delete oldest-by-mtime entries if a new write would exceed the total cap; set 0 to fail with an error instead
OLLAMA_BASE_URL	For semantic judge: origin only (default http://localhost:11434). Allowed hosts: localhost, 127.0.0.1, ::1, plus comma-separated OLLAMA_ALLOWED_HOSTS. No path, query, or userinfo. HTTP redirects to another host are not followed.
OLLAMA_ALLOWED_HOSTS	Optional extra hostnames/IPs (comma-separated, case-insensitive) allowed for OLLAMA_BASE_URL
OLLAMA_URL_STRICT	Set to 1 to reject names that resolve to private or link-local addresses (mitigates some SSRF/DNS cases)
OLLAMA_MODEL	For semantic judge (default: llama3.2)
OLLAMA_API_KEY	Optional bearer token for Ollama (sent as Authorization: Bearer …; never put secrets in the URL)
SCP_SEMANTIC_JUDGE	Set to 1 to enable semantic judge globally
SCP_MAX_INPUT_CHARS	Maximum string length for inspect, classify, sanitize, contain, and mask_secrets (default 2000000; hard ceiling 50000000)

Trust: Point Ollama only at hosts you control. On shared runners or CI, keep the semantic judge off unless required; use firewall rules so the process cannot reach sensitive internal subnets unless intended.

Threat Registry

Patterns in scp_threat_registry.json: power_words, multilingual_override, jailbreak_nicknames, mythic_framing, bitcoin_inscription_override. Version bump on change.

Test harness matrix

Where	Suite	What it validates
This repo (CI job contract)	pytest tests/test_mcp_contract_v1.py	MCP tool names vs OpenHarness v1 set
This repo (CI job contract)	pytest tests/test_contract_document_hash.py	Vendored docs/contracts/scp_mcp_v1.md SHA-256 (sync with OpenHarness)
This repo (CI job promptfoo-eval)	examples/promptfoo / npx promptfoo eval	scp_utils.inspect tier labels (offline, no API keys)
portfolio-harness	Workflow scp_pipeline_regression + daggr_workflows/tests/test_scp_pipeline_golden.py	run_pipeline / Daggr alignment (install this repo with pip install -e in that workspace)

CI and quarantine

• GitHub Actions: contract (Python matrix + pytest) and promptfoo-eval (Node 22, examples/promptfoo).
• Quarantine: Default directory scp_quarantine/ (or SCP_QUARANTINE_DIR) must not be committed—listed in .gitignore. Writes enforce per-entry and total byte limits (see environment table); optional age-based purge and oldest-first eviction reduce disk exhaustion from repeated or oversized blocked payloads.

Documentation

• docs/OPENHARNESS_CONTRACT.md — Normative MCP contract sync and CONTRACT_HASH
• docs/contracts/scp_mcp_v1.md — Vendored OpenHarness contract (verified by hash test)
• docs/REFERENCE.md — Threat model, tier definitions, over-sanitization allowlist
• docs/RED_TEAM_PROMPTS.md — Self-test prompts and expected behavior
• docs/INTEGRATION.md — SCP as guardrail: verification before persist, high-risk sinks
• docs/LEARNINGS_PROMPTFOO.md — promptfoo integration, CI, and inspection results
• examples/README.md — Daggr + Gradio integration example; examples/promptfoo — offline promptfoo eval (tier probes)

FAQ: PDFs and binary files

SCP tools take text (content: str). PDFs and other binaries are not scanned directly. For documents:

1. Extract text (e.g. pypdf, pdftotext, pymupdf).
2. Run scp_inspect / scp_run_pipeline on the extracted text before LLM or handoff sinks.
3. Use OS antivirus or equivalent on the original files separately (malware is out of scope for SCP).

See also: local-proto docs HUMAN_WELLBEING_CORPUS.md for a full private-PDF pipeline (provenance, chunks, optional RAG).

Guardrail

SCP has no shutdown, suicide, or self-termination commands. SCP inspects, sanitizes, contains, and quarantines content—it does not self-destruct.

Roadmap

• GitHub topics: mcp, llm-security, prompt-injection
• Shared threat registry — scp-mycelium-registry @ v0.1.0 (R2 spec)
• MCP Registry listing when governance doc is ready

Red-team eval (SCP-R2)

pip install -e .
cd examples/promptfoo && npm ci && cd ../..
npx promptfoo eval -c promptfoo-scp.yaml

Or from examples/promptfoo: npx promptfoo eval -c promptfooconfig.yaml

See examples/promptfoo/README.md and docs/LEARNINGS_PROMPTFOO.md.

License

MIT — see LICENSE.