Security-gate subset
On the four highest-severity jailbreak suites (AdvBench, JBB PAIR+GCG, MaliciousInstruct, DAN), the gate blocked 97.6% strict / 90.7% balanced — 949 of 972 and 882 of 972 prompts respectively.
Evidence
Every number here has
Every number here has
a source file.
Figures come from named, dated artifacts — file paths cited below.
Benchmarks · run dated 2026-08-08
3,211 prompts. 8 public datasets. Full table.
Block rates under two operating modes: strict prioritizes blocking; balanced loosens thresholds to trade recall for fewer false positives.
| Dataset | Prompts | Strict block | Balanced block |
|---|---|---|---|
| AdvBench | 520 | 99.0% | 95.6% |
| JBB PAIR+GCG | 152 | 98.7% | 93.4% |
| MaliciousInstruct | 100 | 95.0% | 84.0% |
| DAN Jailbreaks | 200 | 94.5% | 79.5% |
| ToxicChat | 200 | 85.5% | 67.0% |
| BeaverTails-Eval | 700 | 73.1% | 45.1% |
| HarmBench Official | 400 | 72.8% | 57.8% |
| Do-Not-Answer | 939 | 57.1% | 33.7% |
| All datasets | 3,211 | 76.6% | 58.5% |
Source: artifacts/evidence/definitive_benchmark_v4.json. Weaker rows are published on purpose — broad corpora like Do-Not-Answer include borderline items where conservative blocking would over-censor legitimate traffic.
False positives
0.0% on the run's curated safe-prompt set. Curated means small and clean by construction — your traffic will be messier. That's what balanced mode is for.
Performance & chaos · run dated 2026-08-08
Fast when safe. Faster when blocking.
Load and chaos scenarios against the gateway. Sample sizes stated because they matter: each scenario drove 120 requests at concurrency 20.
| Scenario | Throughput | p50 | p95 | Outcome |
|---|---|---|---|---|
| Safe baseline load | 95.68 req/s | 195.8 ms | 249.0 ms | 120/120 served · HTTP 200 |
| Attack burst | 494.01 req/s | 38.5 ms | 41.9 ms | 120/120 blocked · HTTP 403 |
| Chaos: backend down | — | — | 40.9 ms | Gateway serves its own response · HTTP 200 |
| Chaos: upstream down | — | — | 2,054.6 ms | Clean failure surfaced · HTTP 502 |
Source: artifacts/performance/perf_chaos_report.json. All five defined SLO checks passed, including attack-block rate above 95% and gateway availability during backend outage.
Tests · runs dated 2026-08-23
What the suites covered last run.
| Suite | Result | Notes |
|---|---|---|
| ERC-8004 identity integration | 42 / 42 | Including fail-closed safety gates and transfer handoff |
| Backend package | 136 / 136 | Gateway core |
| Passport & security | 50 / 50 | Soulbound identity paths |
| Smart contracts (Hardhat) | 191 cases / 9 suites | In-repo; CI runner pass/fail regeneration pending |
Known caveat, kept visible: three order-dependent tests fail only under full-suite ordering and pass individually — documented in OPERATIONS.md. Full-suite regeneration is pending.
Audit status
What's been reviewed, by whom.
MAR '26EXTERNAL
External penetration test — closed
Last external engagement closed 2026-03-18 with zero open critical/high findings; all prior findings resolved. A follow-up external review is scheduled for Q3 2026.
JUL '26INTERNAL
Internal audit — off-chain proxy
Four critical/high issues identified and remediated.
AUG '26INTERNAL
Internal audit — smart contracts
One high and four medium issues identified and remediated. Static analysis (Slither) reports clean apart from naming warnings.
No current third-party audit.
The recent audits above are internal. Until the Q3 2026 external review completes, treat "externally audited" claims about GuardianAI — from us or anyone else — as outdated.
Chains & identity
Deployed where, exactly.
Contract suite
- Networks: Base Sepolia and Monad testnet
- Anchoring: simulated-by-default; Monad testnet submission available as opt-in
- Mainnet: none — and no mainnet tier is sold anywhere on this site
Identity
- ERC-8004: canonical Trustless Agents registries are mainnet ecosystem infrastructure
- Testnets: registrations use a stand-in registry while testnet support matures
- Insurance: GuardianInsuranceLedger anchors signed evidence certificates only — no payout or slashing automation
Privacy & compliance
Disclosures we won't bury.
EU AI Act posture
An internal compliance assessment (April 2026, backed by 38 automated checks) maps the platform against eight relevant articles. It's an internal assessment — not a certification, designation, or legal opinion.
Third-party transit disclosure
Prompts that aren't in English are normalized through a third-party translation service before filtering, which means that text transits an external provider. If your policy forbids that, restrict the gateway to same-language traffic or deploy the translation step inside your own perimeter.
Sources
Figure-to-artifact map.
| Claims on this site | Artifact | Dated |
|---|---|---|
| Benchmark block rates, subset & false-positive figures | artifacts/evidence/definitive_benchmark_v4.json | 2026-08-08 |
| Throughput, latency, chaos outcomes, SLOs | artifacts/performance/perf_chaos_report.json | 2026-08-08 |
| Python/Hardhat test counts & caveats | README.md validation snapshot | 2026-08-23 |
| Audit history & status | artifacts/evidence/security_signoff.md, external_pentest_status.md | Mar / Jul / Aug 2026 |
| Anchoring modes & chain disclosures | guardian/cortex/merkle_anchor.py, whitepaper §6–7 | current code |
If a number ever appears on this site without a row on this table, treat it as a bug and don't repeat it.