Move over, Mythos. Here comes... pretty much any other model with a good harness

What it is

An Aikido blog post arguing that for AI vulnerability discovery, the orchestration harness matters more than the frontier model inside it. The concrete, actionable payload is a named reference harness architecture (from Cloudflare's research) plus specific evidence that models are interchangeable inside a good harness.

How it works

The Cloudflare-style pipeline the post recommends has five distinct stages:

1. Recon — reads the repo and builds a task queue for everything downstream.
2. Hunt — many agents run in parallel, each searching for vulnerabilities (relying on LLM non-determinism so different agents surface different bugs).
3. Validation — an independent agent with a different prompt and no ability to generate its own findings tries to disprove what the hunt agents found.
4. Tracing — follows confirmed findings to determine whether attacker-controlled input can actually reach the bug from outside.
5. Deduplication — consolidates findings sharing a root cause.

The separation of "hunt" from "validation" (validator cannot author findings, only refute) is the specific design detail worth copying — it directly attacks the false-positive problem the cURL incident illustrates.

What to do

• Invest in the harness, not model selection. Two data points quantify why: UCSB found the same Claude Opus 4.6 passed 4× as many tests under the best vs. worst harness, while frontier models (Opus 4.6 vs GPT-5.4) differ by ~1pp on coding benchmarks.
• Design for model-swapping. Mozilla reports that once their pipeline was solid, each new model dropped in improved bug-finding with no rearchitecting. Niels Provos built a harness that found an 18-year-old vuln, then swapped in open-weight GLM 5.1 and got comparable results — so a strong harness can use cheaper/open models.

Caveats / extra facts the summary omits

• Frontier gap is small and non-directional: UK AISI benchmarked GPT-5.5 at 71.4% vs Mythos 68.6% on the hardest category; "one beats the other depending on the task."
• Public Mythos is unusable for this work: Fable 5 (the public release) is Mythos 5 with guardrails that halt on any cybersecurity or biology request — it cannot be benchmarked or used for vuln finding at all. Plan around API/enterprise access.
• Cost is a real constraint: the post states running Mythos once costs "around tens of thousands of dollars" for a single run (text cuts off mid-sentence), reinforcing the case for cheaper swappable models under a good harness.
• The cURL/Stenberg example (3 false positives, 1 non-security bug, 1 real vuln out of 5 "confirmed") is the concrete argument for the mandatory independent-validation stage above.