Repository navigation
Releases
v0.2.1 — the number the model wrote
A hunt died on the number the model wrote. Run end to end against a deliberately vulnerable application, with a local model: it answered "repeat": 3, SequencePlan refuses a count below five, RunAudit passed the model's number straight in, and nothing caught the refusal. The whole hunt ended on an uncaught exception, exit 255, after several minutes of inference — every finding lost.
Three hundred and ten unit tests passed through that, because none of them ever handed the domain a number a model had chosen. The count is input like the rest of the plan, so it is brought inside the bounds where the rest of the plan is corrected. The constructor stays strict for anyone calling it with a number of their own, and the Next.js twin already bounded it this way.
v0.2.0 — everything the twin learned
Everything the Next.js twin learned, brought across. The two tools are the same tool for two frameworks, and the Symfony one had fallen a week behind. Every defect below was found on the other side, verified to exist here, and fixed with the tests that prove it.
A refusal is not the flaw. A plan was refused only when every status it recognised was a refusal, so one sitting next to a real one went through — and the runner asks whether the answer is in the list, which turns that refusal into a proof. Measured on a live run: a model answered [200, 404] for a path the application does not serve, the server said 404, and the report carried a proven finding whose entire evidence was the page not existing.
A read that deletes a row is not a read. The safety policy refused anything but GET, HEAD and OPTIONS, then let through a GET carrying X-HTTP-Method-Override: DELETE. Symfony honours that header, and ?_method= once enableHttpMethodParameterOverride() is on — one line in a front controller, and plenty of applications have it. The method that will actually be routed decides the verdict now.
--local keeps its promise. The mode says nothing leaves the machine, then read OLLAMA_HOST without looking at it. The address goes through the same check as an attack target, and the form Ollama documents — localhost:11434, without a scheme — is understood instead of failing on a parser error.
The test written for the team was writing itself wrong. It asserted on the status whatever the proof was, so a leak found by reading the body was answered 200 by a sound application too and the emitted test failed forever. Several statuses were interpolated into assertNotSame, which took the second one as the value under test and the response as the failure message: the test then passed whatever the application answered. And everything in that file comes from a model or from a hand-edited baseline — a quote in a path closed the call and opened a statement of its own. It is input now, and it is escaped like input.
Smaller, and still worth saying. A baseline whose shape is wrong is refused instead of read as empty, naming the entry so the file is easy to fix. A path that is not there is named instead of the project being called clean. A SARIF location is written as a uri, so a space or an accent in a file name stops breaking the viewer. The report is pinned to English rather than following the operator's own settings, since the same repository must not read differently depending on who scanned it.
And the page no longer contradicts itself about what a repeated request proves.
v0.1.5 — one line per hole
Findings that share a rule, a file and a request are collapsed into the worst of them.
Measured on a deliberately sabotaged copy of a real Symfony starter, the same run went from six proven findings to four: four of them described the same unguarded route from different angles. A reader who sees four holes fixes the first and finds three still there the next day.
Two different rules stay apart, because each one needs its own fix. A finding with no request behind it is never collapsed, since nothing says two of them are the same.
The action no longer requires an API key: a runner needs a key or a model of its own, reachable through OLLAMA_HOST or OPENAI_BASE_URL. It also no longer carries the previous name it shared with another project.
v0.1.4 — one line per hole
Findings that share a rule, a file and a request are collapsed into the worst of them.
Measured on a sabotaged copy of a real starter application, the same run went from six proven findings to four: four of them described the same unguarded route from different angles. A reader who sees four holes fixes the first and finds three still there tomorrow.
Two different rules stay apart, because each one needs its own fix. A finding with no request is never collapsed, since nothing says two of them are the same.