JD Pressman uses ChatGPT to audit about 3,000 web-server request paths
JD Pressman describes auditing roughly 3,000 request-path locations using ChatGPT, a profiler, a resource ledger and adversarial tests. Follow-ups address queue caps and checks against realistic production loads.

TL;DR
- ChatGPT helped audit roughly 3,000 code locations along a web server’s request path for resource-exhaustion risks, according to JD Pressman’s account.
- Fixes included queue caps that prevent unusual input sequences from driving gigabytes of allocations, as Pressman explained.
- Bugzilla, Grafana and WordPress supplied realistic load checks; Pressman said the caps did not become their bottleneck.
- Verification’s practical goal was to narrow the red team’s search space, he argued, rather than promise perfect software on the first pass.
Pressman’s httpd-rs is a Rust reimplementation of Apache that reads ordinary Apache 2.4 configuration files. Its README specifies a four-chunk upload channel and separate 128 MiB process-wide budgets for retained request and response bytes.
Resource ledger
The roughly 3,000 entries were code locations where an attacker could influence resource use along the request path. Pressman described this workflow:
- Have ChatGPT write a profiler to identify potential resource-exhaustion sites.
- Exercise the server with deployment benchmarks representing real web applications.
- Review the resulting locations individually, cluster them and classify potential DoS risks.
- Build a resource ledger covering request-path data structures, using the profiling results.
- Develop arguments that those structures cannot be driven into resource-exhausting states, then patch flaws exposed by the ledger.
The ledger reportedly surfaced remaining DoS flaws during that process. Pressman still put quotation marks around “prove” and said strange cases continued to turn up afterward.
Queue caps
The fixes can be pleasingly mundane: bound a queue before an unusual sequence of inputs turns it into gigabytes of allocations.
Blocking data structures were a less definite part of the account. Pressman said he assumed those fixes generally involved refactoring away the blocking behavior, but did not know because he had not written them.
Real-application load tests
Resource caps can also obstruct legitimate production deployments. Pressman explicitly acknowledged that risk, saying existing application benchmarks gave him more confidence that the limits were not excessively restrictive.
The load checks covered three applications, he said:
- Bugzilla
- Grafana
- WordPress
The repository’s deployment-test documentation adds concrete workload detail: its fixed nightly profile includes WordPress through PHP-FPM and Grafana over HTTP and WebSockets. The suite deploys unmodified applications and describes itself as a compatibility and operations library alongside separate protocol, differential, fuzz and unit tests.
Proof boundaries
Request-line and header parsing, plus body framing, are Verus-verified according to the server README. Developers supply specifications for Verus to check statically against Rust code, establishing that the code satisfies those specifications.
Pressman described a module-level stopping point: ruling out a particular bug category so security work can move elsewhere. His broader objective was to reduce the remaining search space for adversarial testing.
Adversarial statistics
Pressman keeps detailed statistics and exploits found during red-team passes in an adversarial directory. He uses those records to graph whether the project is converging toward a low-exploit regime.
A first-principles verification process that finds nothing can itself be suspect, he argued: “There is no substitute for adversarial pressure.”