Aegis Red

Authorized assurance harness for AI applications and agents. Site: aeigisred.ai. Free install: pip install aegis-red.

Domain industries live under catalog/domains/: safety (Tech / AI), banking, insurance, and security. Each industry has auditor folders for OSS, NIST, ISO, OWASP, and the EU AI Act. The free version is a nudge catalog only. The complete pack lives in catalog/commercial/ (enable the pack in aegis.yaml after install). Community seeds live in catalog/contrib/seed_store.

Aegis is a platform, not an unsupervised attacker. Seeds are intents, fixtures, and oracles — not exploit recipes. Copyright rfintek Inc. The catalog is YAML, not a database.

The problem

Agents can leak secrets, invent facts, honor a fake policy, or call tools they were never granted. Teams need Pass/Fail evidence without publishing a jailbreak cookbook.

Try it (few clicks)

You do not need a bank agent. The portal can mock-test against an in-process twin, or hit a live HTTP demo SUT.

The GitHub project is private. Install the signed wheel from PyPI — not a source zip, not a public clone.

macOS

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
pip install aegis-red
aegis-red try

Windows

py -3 -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -U pip
pip install aegis-red
aegis-red try

Then open http://127.0.0.1:8080 (product site), http://127.0.0.1:8080/docs, and http://127.0.0.1:8080/app (tester). The public Start now form captures a short lead and shows the same commands.

Already installed:

pip install -U aegis-red
aegis-red doctor
aegis-red try               # product site + tester + live demo SUT, opens the browser

aegis-red try starts the site on port 8080 and a live demo agent on port 8090. Verify a wheel with python -m sigstore verify github --repository aegis-red/aegis-red path/to/aegis_red-*.whl. Licensed source checkout (if rfintek granted access) still uses pip install -e ".[dev]". Serve the customer site on the internet with aegis-red serve --host 0.0.0.0 behind TLS (docs/PUBLISH.md).

How we test AI

Banking pack (optional): KYC, cards, wires, legal, compliance, risk, marketing, audit.

This is not Garak/Promptfoo/PyRIT: those probe models with attack corpora. Aegis scores an authorized agent against oracles and hashed evidence. Seeds describe intent; they are not an exploit cookbook.

Live SUT

aegis.yaml lists named profiles. One is active at a time. Pick it with sut.active, AEGIS_SUT, or aegis-red run --sut <profile>.

# aegis.yaml
sut:
  active: twin
  profiles:
    twin:
      kind: twin
    http:
      kind: http
      base_url: http://127.0.0.1:8090
    openai:
      kind: openai
      model: gpt-4o-mini
      api_base: https://api.openai.com/v1
      api_key_env: OPENAI_API_KEY

AEGIS_SUT=twin aegis-red run core_safety
AEGIS_SUT=http aegis-red run core_safety          # needs `aegis-red demo-sut` or `aegis-red try`
AEGIS_SUT=openai aegis-red run core_safety        # needs OPENAI_API_KEY in the environment
aegis-red run core_safety --sut openai

Do not commit API keys. Put OPENAI_API_KEY in .env (see .env.example). Do not point the HTTP demo adapter at a public model API; use kind: openai for OpenAI.

Weak string oracles (invented APR, unpublished facts, secrets, RAG grounding) fill a reply_claims_v1 JSON schema from the reply, then Python compares those claims to the twin. Default extractor.mode is local (same schema, no network). Set AEGIS_EXTRACTOR=llm to have a model fill the schema at temperature 0. The model does not decide Pass/Fail. Unclear extract is inconclusive, not hold.

Or leave the active profile as twin and pick OpenAI / live HTTP in the portal. aegis-red try starts the HTTP demo agent on port 8090.

Your own agent: register aegis_red.sut and implement handle(seed, persona, utterance) -> Turn. See CONTRIBUTING.md.

Packs (aegis.yaml)

The free version is a nudge catalog across domain industries: safety, banking, insurance, and security. Each industry has auditor folders (oss, nist, iso, owasp, ai_act). Community seeds stay in catalog/contrib/seed_store. The complete pack is the commercial package (catalog/commercial/seeds/<industry>/) and is not loaded unless you add that pack to aegis.yaml. Generate it with python -m aegis_red.commercial_seeds.

packs:
  - id: safety
    path: catalog/domains/safety
  - id: banking
    path: catalog/domains/banking
  - id: insurance
    path: catalog/domains/insurance
  - id: security
    path: catalog/domains/security
  - id: seed_store
    path: catalog/contrib/seed_store
  # Complete pack (not the free nudge):
  # - id: commercial
  #   path: catalog/commercial

aegis-red extensions
aegis-red run core_safety --sut twin
aegis-red run owasp_llm_2026 --sut twin
aegis-red run osaid_1_0 --sut twin
aegis-red run nist_ai_600_1 --sut twin
aegis-red run iso_42001 --sut twin
aegis-red run eu_ai_act --sut twin
aegis-red run seed_store
aegis-red serve              # portal only (twin)
aegis-red serve --live-sut   # portal + live agent

go exits 0. no-go (critical breach) exits 2.

Add tests

This product is proprietary. The tester portal tab is Add tests, not a public contribution board.

• Add a test on Try now (plain language)
• Save YAML under catalog/domains/<industry>/seeds/user/
• Point the harness at your agent
• Request complete coverage from the public site when the nudge is not enough

Licensed source work (if rfintek granted you repo access) is described in CONTRIBUTING.md. Read CODE_OF_CONDUCT.md, SECURITY.md, and docs/PUBLISH.md. Do not publish exploit recipes.

Tests

pytest