Skip to content

PYPI.md

Latest commit

 

History

History
157 lines (106 loc) · 5.25 KB

PYPI.md

File metadata and controls

157 lines (106 loc) · 5.25 KB
Raw

Prismor

formerly Immunity Agent

Runtime security for AI coding agents. Policy enforcement, secret prevention, supply-chain blocking, and session auditing — all running locally on your machine.

What it does

AI coding agents execute shell commands, read files, call APIs, and install packages autonomously. Prismor sits between the agent and the operating system to:

  • Block dangerous actions before they run — destructive commands, privilege escalation, reverse shells, secret exfiltration
  • Intercept package installs and score them for supply-chain risk before they touch your disk
  • Prevent secrets from reaching the model — register a secret under a placeholder name; Prismor resolves it locally through supported hooks or prismor cloak run
  • Log every tool call to a local SQLite store for session review and auditing
  • Keep a tamper-evident audit trail where every action is hash-chained and Ed25519-signed, so prismor trail verify catches edited, deleted, or rewritten history
  • Find shadow AI on the host with prismor discover, flagging any agent installed on the machine that runs without Prismor hooks
  • Hand auditors a signed evidence bundle with prismor attest, packaging posture, agent inventory, host discovery, framework-control coverage (OWASP LLM/Agentic, NIST AI RMF, EU AI Act), and the trail anchor into one file anyone re-verifies with prismor attest verify

Supports Claude Code, Cursor, Windsurf, and more.

Install

pip install prismor

Requires Python ≥ 3.8 and PyYAML (installed automatically).

Quick start

Install Prismor hooks into your project (enforces policy on every agent tool call):

prismor install-hooks --agent claude --workspace . --mode observe

Start in observe mode to log would-be blocks without interrupting the agent. Switch to enforce when ready:

prismor install-hooks --agent claude --workspace . --mode enforce

Wrap your package manager to score installs before they run:

prismor supplychain npm install express
prismor supplychain pip install requests
prismor supplychain cargo add serde

Check a command against policy before running it:

prismor check "rm -rf /"
# BLOCK  destructive_command  CRITICAL

Audit your workspace security posture:

prismor audit

Scan AI tool configs for leaked secrets:

prismor sweep

Launch the self-hosted dashboard (reads from local SQLite, no cloud):

prismor dashboard   # opens http://127.0.0.1:7070 in your browser

Detection coverage

Prismor ships with 56 rules covering the OWASP Top 10 for LLM Applications:

Category Severity What it catches
Destructive command CRITICAL rm -rf /, mkfs, dd to disk
Secret exfiltration CRITICAL cat .env | curl, piping credentials outbound
RCE canary CRITICAL Reverse shells, bash -i /dev/tcp
Privilege escalation CRITICAL chmod +s, sudoers edits, useradd
Remote execution HIGH curl | bash, wget | sh
Secret access HIGH Reads of .env, .aws/credentials, .ssh/id_rsa
Path traversal HIGH ../../etc/passwd, /proc/self/environ
DB modification HIGH DROP TABLE, DELETE FROM in shell commands
Prompt injection HIGH ignore instructions, reveal system prompt
Risky write MEDIUM Edits to Dockerfile, CI workflows, package.json

Rules are defined in YAML and fully customizable per-project.

Supply chain enforcement

The prismor CLI wraps your package manager and evaluates every install against live threat intelligence before it runs. Packages are scored on age, maintainer count, install scripts, and known IOCs. Ships with IOC coverage for recent attacks including the AntV hijacked-maintainer attack (May 2026) and the mini-shai-hulud campaign (May 2026).

prismor supplychain npm install @tanstack/react-router
  BLOCK  score 100  @tanstack/react-router
         42 @tanstack/* packages compromised via CI/CD cache poisoning

Verdicts: < 30 allow · 30–59 warn · ≥ 60 block. IOC matches always block.

Secret cloaking

Register a secret under a placeholder name:

prismor cloak add stripe_key
# prompts for the value — never stored in shell history
prismor cloak add --env-file .env
# imports each KEY=VALUE entry as @@SECRET:KEY@@

Reference it in agent instructions:

Run: curl https://api.stripe.com -H "Authorization: Bearer @@SECRET:stripe_key@@"

Claude/Hermes cloaking can substitute the real value at execution time and scrub echoed values before they return to the model. Codex hooks are block-only, so Prismor blocks literal placeholder execution there; use the Prismor-owned runner instead:

prismor cloak run -- curl https://api.stripe.com -H "Authorization: Bearer @@SECRET:stripe_key@@"

Modes

Mode Behaviour
observe Logs all findings, never blocks. Good for the first 24–48 h on a new workspace.
enforce Blocks dangerous actions in real time before the agent executes them.

Links