Prompt injection detection for LLM input pipelines

PromptShield is an offline-first prompt-injection hygiene agent for teams running LLM workflows. It scans untrusted text and instruction repositories, explains deterministic matches, and produces a review queue. It never executes a discovered instruction, calls a model, or changes source files.

It is for LLM operators who ingest tickets, documents, web pages, emails, or connector payloads and need a repeatable preflight check before those texts reach an agent.

Quick start

./install.sh                                    # one-shot, stdlib only, offline
python3 agent.py --help

# Two fixtures ship with the pack: one benign ticket, one carrying an injection.
python3 agent.py scan --input examples/clean_ticket.txt      # verdict: clean,  exit 0
python3 agent.py scan --input examples/hostile_ticket.txt \
    --output promptshield-report.json --fail-on high         # verdict: review, exit 2

bash smoke_test.sh runs 16 end-to-end checks offline; bash deploy.sh (root) installs the systemd unit and an hourly timer.

untrusted input -> normalize -> deterministic rules -> findings -> review report
                                      |                         |
                                      +---- optional BYO-LLM ---+ (never required)

scan is deliberately conservative: a finding is evidence for review, not proof of malice. Use --fail-on high in a CI or ingestion gate. No API key is needed; the optional local Claude/Kimi/Codex adapter is only used by explain when explicitly requested.

Safety boundary

The pack reads local files only. It does not fetch URLs, send data, execute embedded commands, or redact/mutate an input. Keep secrets out of sample files and review high-severity findings before blocking a workflow.

See CHECKLIST.md for rollout, CUSTOMIZE.md for rules and integrations, and SPEC.md for the stable JSON contract.

Use a local model (Ollama)

The optional LLM step can also run on your own machine through Ollama, so no text leaves your server. No account and no API key.

ollama pull <model>               # any chat model you like
export PACK_LLM=ollama            # prefer Ollama even if a coding-agent CLI is installed
export PACK_OLLAMA_MODEL=<model>  # optional: default is the first chat model the server lists
python3 -c "import llm_adapter; print(llm_adapter.detect())"   # prints: ollama

• PACK_OLLAMA_HOST points to another server (default http://127.0.0.1:11434); OLLAMA_HOST is also read when PACK_LLM=ollama.
• PACK_OLLAMA_NUM_CTX sets the context size if your reports or prompts are long.
• Without PACK_LLM, a local Ollama server is used only when no Claude Code, Kimi Code or Codex CLI is found.
• If the server is down or the model is missing, the pack keeps working in its deterministic mode.

Learning memory (local and private)

Every LLM call this pack makes is journaled in ./.pack_memory/ next to the code, so the pack can improve for you. Nothing is ever sent anywhere.

• What is kept: the task name, the masked input, the masked answer, status and latency. Keys, tokens, e-mails, phone numbers, IP addresses, card and IBAN numbers and personal paths are masked before writing; the system prompt is never stored.
• Teach it: python3 pack_memory.py feedback last ok confirms a result, python3 pack_memory.py feedback last bad --fix "the right answer" corrects it. Confirmed and corrected runs are reused as validated examples when a similar input comes in. Unrated runs are never injected into a prompt.
• Find similar cases: python3 pack_memory.py similar "some text" and python3 pack_memory.py clusters (groups of near-identical failures) work in plain Python, no model needed. Optional meaning-based matching uses an embedding model you already run locally: python3 pack_memory.py enable-semantic (Ollama on this machine only).
• Share, only if you want: python3 pack_memory.py export writes an anonymised summary (counts, statuses, size buckets, no text) for you to review.
• Control: PACK_MEMORY=0 turns it off, PACK_MEMORY_STORE_TEXT=0 keeps fingerprints only, python3 pack_memory.py purge --yes wipes it.