The manifesto.

Nobody owns the seam between offensive-AI and AI-SOC: detection coverage drift, validated by real exploits.

The wedge.

There are two crowded rooms in AI security right now. In one, offensive-AI tools race to demonstrate more end-to-end attack capability: how far an agent can get, how many steps it can chain, how impressive the kill chain looks in a demo. In the other, AI-SOC tools race to triage alerts faster, summarize incidents, and write detections at the speed of a language model.

Between those two rooms there is a seam nobody owns: detection coverage drift, validated by real exploits. Your attackers change. Your stack changes. The detections you wrote last quarter quietly stop matching what an adversary actually does. The offensive room proves attacks but does not ship you a detection. The defensive room ships detections but cannot prove they fire on a real exploit. Riposte lives in that seam, and the seam is the whole point.

Generation is the easy part.

A language model will happily write you a Sigma rule. It will be well-formed, plausibly named, and confidently wrong. LLM-written rules are often syntactically perfect and logically wrong, so the verification harness, not the generation, is the product.

A rule that compiles is not a rule that works. A rule that works on the exact payload you fed the model is not a rule that survives a mutation. A rule that catches the attack is not a rule you can run in production without drowning your analysts in false positives. Each of those is a separate question, and a generator answers none of them. The harness answers all four, and it answers them by failing closed: a detection ships only if it compiles, fires on the real exploit, survives a model-mutated variant, and stays under your false-positive threshold.

Why it is called Riposte.

In fencing, a riposte is the counter-attack you make immediately after parrying your opponent's blade. The parry and the riposte are one motion: you do not first defend and then, in a separate beat, attack. The defense is what creates the opening for the offense.

That is the shape of the product. Red reproduces an exploit on an authorized target. Blue drafts a detection from that exact reproducible artifact. The offense and the defense are one motion, and the detection you ship is the direct consequence of the attack you proved. Every exploit your AI proves becomes a detection your SIEM did not have.

Honesty is the brand.

Riposte is pre-alpha. There is no hosted product, no published benchmark numbers, no third-party engagement. It has not run unsupervised, it has not run against a third party, and it has not shipped a verified detection outside a self-owned lab. The site you are reading practices the product's own ethos: no third-party trackers, self-hosted fonts, minimal JavaScript, no chat widget, no fake logos, no testimonials.

A security tool that overclaims is worse than useless, because the one thing it sells is trust. So the rule here is simple: every claim survives a git clone. If the site says the harness holds bad rules back, you can run the demo and watch it hold bad rules back. If it says there is no telemetry, you can read the code and find none. The harness is built to catch rules that are confidently wrong. The least we can do is hold the marketing to the same standard.