Multi-protocol compliance validation with dual-agent cross-review.
AI coding tools generate infrastructure faster than anyone can validate it. The result: DNS misconfigurations, expired certificates, weak cipher suites, and CA/Browser Forum violations that slip into production undetected.
Existing tools are reactive (scan after deploy), single-authority (one engine, one opinion), and passive (run once, forget). None of them separate the security verdict from the compliance verdict — so when they disagree, you never know.
Bellerophon probes targets across protocol layers, evaluates the collected evidence against declarative OPA Rego policies, and detects what changed between checks. Two independent agents — one for security, one for compliance — run against the same evidence and cross-review each other's findings to produce a confidence-scored verdict.
Agreement means confidence. Disagreement means a finding.
In Greek mythology, Bellerophon rode the winged horse Pegasus to slay the Chimera — a fire-breathing hybrid that no single hero could face alone.
The names are not decoration. They map directly to the architecture:
- Bellerophon (security agent) — asks: "Can this be exploited?" Evaluates
security/andshared/policies. Thinks like an adversary. - Pegasus (compliance agent) — asks: "Does this meet the standard?" Evaluates
compliance/andshared/policies. Validates against CA/Browser Forum Baseline Requirements and industry standards. - Chimera — the AI-generated code security platform that Bellerophon was built to validate. First consumer of these compliance reports.
When Bellerophon and Pegasus agree, you have a high-confidence verdict. When they disagree, the cross-review flags the divergence as a finding. Neither agent can self-approve.
# Clone and install
git clone https://github.com/Guipetris/bellerophon && cd bellerophon
cargo install --path .
# First check (text output, all policies)
bellerophon check https://your-target.com
# JSON output for scripting
bellerophon check https://your-target.com -f json
# Dual-agent cross-review (highest confidence)
bellerophon review https://your-target.comDefault --policy-dir is ./policies. Default --store-dir is ./evidence-store.
| Layer | What's Extracted | Status |
|---|---|---|
| DNS | IP resolution, endpoint discovery | Active |
| TLS | Certificate chain, key types/sizes, protocol version, cipher suite | Active |
| HTTP | Security headers (HSTS, CSP, X-Frame-Options), info leakage | Coming Soon |
| OIDC | Discovery document, signing algorithms, grant types | Coming Soon |
| DNS CAA | Certificate Authority Authorization records | Coming Soon |
| Policy | Agent | What It Checks |
|---|---|---|
shared/tls/certificate_chain_valid |
Both | Chain has leaf + CA, links correctly, terminates at CA |
shared/tls/certificate_not_expired |
Both | No certificate in chain is expired or approaching expiry |
shared/tls/minimum_key_strength |
Both | RSA >= 2048 bits, ECDSA >= 256 bits |
security/tls/weak_cipher_detection |
Bellerophon | Protocol >= TLS 1.2, strong cipher suite, adequate key exchange |
compliance/tls/cabf_baseline |
Pegasus | CA/Browser Forum Baseline Requirements (validity period, key size, signature algorithm) |
| Policy | Awaiting |
|---|---|
dns/caa_records_present |
DNS CAA probe |
http/info_leakage |
HTTP probe |
http/security_headers_present |
HTTP probe |
oidc/discovery_present |
OIDC probe |
oidc/secure_token_config |
OIDC probe |
Stub policies are loaded but produce SKIP results when their required evidence type has not been collected. They do not fail the scan.
| Format | Flag | Status | Notes |
|---|---|---|---|
| Text | -f text |
Implemented | Colored terminal output with PASS/FAIL/WARN/SKIP/ERROR badges. Default. |
| JSON | -f json |
Implemented | Machine-readable. Includes full ComplianceReport + ChangeReport. |
| SARIF | — | Planned | GitHub Security tab integration. |
| HTML | — | Planned | Customer-facing branded report. Generator exists; not yet wired to CLI. |
| Command | What Runs | Use When |
|---|---|---|
bellerophon scan <TARGET> |
Bellerophon agent (security/ + shared/) |
You want a security-focused assessment |
bellerophon audit <TARGET> |
Pegasus agent (compliance/ + shared/) |
You want a compliance-focused assessment |
bellerophon review <TARGET> |
Both agents independently, then cross-review with confidence score | You want the highest-confidence verdict |
bellerophon check <TARGET> |
All policies (security/ + compliance/ + shared/ + stubs) |
Backward-compatible full scan |
Each subcommand accepts the same flags: -p/--policy-dir, -s/--store-dir, -f/--format.
Run it twice and Bellerophon tells you what changed.
On the second run against the same target, with the same --store-dir, the tool compares the new results against the stored report from the previous check:
Changes since last check (2026-03-21T14:30:00Z)
[REGRESSION ] shared/tls/certificate_not_expired: pass -> fail — certificate expires in 3 days
[IMPROVED ] security/tls/weak_cipher_detection: fail -> pass — protocol upgraded to TLS 1.3
[NEW ] shared/tls/minimum_key_strength: pass — first evaluation
Regressions: 1 | Improvements: 1
Change types: REGRESSION (pass to fail), IMPROVED (fail to pass), NEW (first evaluation), CHANGED (status shift), UNCHANGED.
Change detection requires a consistent --store-dir between runs. The default is ./evidence-store relative to the working directory.
| Code | Meaning | Use In CI |
|---|---|---|
| 0 | Compliant | All evaluated policies passed |
| 1 | Non-compliant | One or more policies failed |
| 2 | Degraded | Warnings only, no hard failures |
| 3 | Error | Probe failed, invalid target, or policy load error |
Exit code 3 is used for infrastructure errors. Exit code 1 is the gating signal for deployment pipelines.
bellerophon check https://target.com || exit 1Bellerophon is being built toward three capabilities:
Discovery (current) — Probe live targets, evaluate against policies, detect drift between checks. Give engineers a compliance posture snapshot they can run in CI.
Certification (next) — Package compliance reports as W3C Verifiable Credentials (SD-JWT-VC format). A signed, cryptographically verifiable attestation that a target passed a specific set of policies at a specific point in time.
Alerting (planned) — Watch mode with continuous monitoring. Detect regressions the moment they appear — certificate near-expiry, cipher suite downgrade, chain validity failure — and emit structured alerts before they reach production.
Planned features, in approximate order:
- HTTP security headers probe (HSTS, CSP, X-Frame-Options, info leakage detection)
- OIDC discovery probe (signing algorithms, grant types, discovery document validation)
- DNS CAA probe (Certificate Authority Authorization record validation)
- SARIF v2.1.0 output (GitHub Security tab integration)
- HTML report output (customer-facing branded report; generator exists, CLI wiring pending)
- Watch mode (
bellerophon watch) for continuous monitoring - S3 evidence store backend (object lock, DynamoDB index, KMS encryption)
- W3C Verifiable Credentials output for compliance certification
Bellerophon probes and policies are mapped to industry threat frameworks and standards:
| Framework | Coverage | Reference |
|---|---|---|
| MITRE ATLAS (AI/ML threats) | 13 AI probes mapped to ATLAS techniques | docs/mitre-atlas-mapping.md |
| MITRE ATT&CK (infrastructure) | 9 infrastructure probes mapped to ATT&CK | docs/mitre-atlas-mapping.md |
| OWASP Top 10 for LLMs | Certification profile with 10 controls | Certification profile: owasp-llm-top10 |
| OWASP ASVS v4 | 11 controls (HTTP, TLS, CORS) | Certification profile: owasp-asvs |
| EU AI Act | Article-level control mapping | Certification profile: eu-ai-act |
| ISO/IEC 42001 | AI management system controls | Certification profile: iso-42001 |
| NIST AI RMF | Govern/Map/Measure/Manage functions | Certification profile: nist-ai-rmf |
OWASP project application: docs/owasp-project-application.md
Chimera is a trust-as-a-service platform for AI-generated code security. It uses weighted voting across multiple validation agents to assign a trust score to generated artifacts. Bellerophon is the forward-edge discovery and validation layer: it probes the infrastructure that Chimera and its consumers depend on, and its compliance reports feed into Chimera's trust signal pipeline. Bellerophon validates Chimera; Chimera consumes the reports. The relationship is not circular — Bellerophon has no dependency on Chimera at runtime. See docs/chimera-integration.md for the planned integration architecture.