Klue

Hydra is the configuration framework a large part of the Python machine-learning world uses to wire up experiments and apps. We pointed KLUE, our AI security agent, at it, and it found two ways to run code on a machine that loads an untrusted Hydra config. Both get around the blocklist Hydra had already added to stop exactly this. One slips through gaps in that blocklist; the other goes through the logging system, a door the blocklist never guarded. Both were reported, accepted, assigned CVEs (CVE-2026-106441 and CVE-2026-106442), rated High, and fixed.

axios is the HTTP client most JavaScript and Node apps use to make requests. While reading through it, our security agent KLUE found a bug in how axios chooses the request method. On the default axios object, if another flaw in your app has polluted Object.prototype, a call you wrote as a safe GET can quietly go out as a DELETE, POST, PUT, or PATCH. Here is how KLUE found it, why only some ways of calling axios are affected, and how it was fixed.

We asked KLUE, our AI security agent, to review kin-openapi, a library that a large share of Go web APIs use to read and check incoming requests. It found three separate bugs that each let anyone crash the service with a single request. The pattern behind all three is the same: the library already had the right safety check, it was just missing on one of the paths that needed it. All three were reported, accepted, assigned CVEs (CVE-2026-100352, CVE-2026-100353, CVE-2026-100354), rated High, and fixed by the maintainers.

A detailed bug bounty writeup. Our autonomous agent KLUE spent hours proving a hardened crypto-mining marketplace was hardened, then reverse-engineered its HMAC signing scheme, mapped a STOMP WebSocket buried in the authenticated app bundles, and discovered it performed no authentication at all, routing private per-tenant channels off a client-supplied parameter. An unauthenticated, cross-tenant data leak, responsibly disclosed and rewarded.

We ran KLUE against the same live production target four times, once each behind GLM 5.2, DeepSeek V4 Pro, Kimi K2.7 and Opus 4.8, all on an identical thirty minute budget. Same target, same clock, four risk verdicts from Low to Critical. A field report on what each model sees, what it walks past, whether it got the severity right, and why the cheapest run came back with the most.