---
title: 'In the News: November 15, 2026 (Morning)'
description: 'NUS researchers show agent-written proofs built on fuzz harnesses can hold while the library stays buggy, and release harnesses that cover the whole API.'
canonical_url: 'https://darkfactory.dev/news/2026-11-15-morning'
markdown_url: 'https://darkfactory.dev/news/2026-11-15-morning.md'
collection: news
date_published: '2026-11-15T07:20:00-05:00'
date_modified: '2026-11-15T07:20:00-05:00'
---

# In the News: November 15, 2026 (Morning)


One item this morning. A group at the National University of Singapore argues that the large agent-written proofs published last month verify less than they appear to, and has released the specifications it thinks agents should be proving instead.

## 1. A proof built on a fuzz harness covers one call sequence, not the library

**[Pre-QED: Environment Specifications for Agentic Verification](https://arxiv.org/abs/2611.15207)** · Tristan Lim, Haoxin Tu, Mehtab Zafar, Zhengxiong Luo, Abhik Roychoudhury, National University of Singapore · arXiv, November 11, 2026

OSS-Verify, from Anthropic's Frontier Red Team and Theorem, takes the fuzz harness an open-source project already keeps for OSS-Fuzz, turns it into the theorem that the program never crashes, and has a Claude agent prove that theorem in Rocq. The paper credits it with proofs across 239 entry points in 123 libraries. The NUS authors point out that a fuzz harness usually makes one fixed sequence of API calls, so the proof covers that sequence. Their worked example is a small stream API whose harness opens a handle, reads once and closes it. The proof holds for every input up to 64 MB, yet any client that reads twice from the same handle triggers an out-of-bounds read: two 40-byte reads on a 64-byte stream run 16 bytes past the end.

Their fix is a "verification harness" that decodes its input into any sequence of calls the API contract allows, with call-order rules written in linear temporal logic and extracted by an agent workflow from the code and documentation. They built harnesses for 29 of the 32 libraries with public OSS-Verify proofs. By their count, average API coverage rises from 22.1% for the fuzz harnesses to 89.3%, and full coverage from 2 libraries to 11. The workflow extracted 234 temporal properties. The authors say "the bottleneck of agentic verification has shifted from how to prove to what to prove." These harnesses have not been proved yet; the team releases them at [pre-qed.github.io](https://pre-qed.github.io/) as open challenges and says proving them is harder, because the prover has to reason about state that persists across arbitrarily long call sequences. They also report that fuzzing with the new harnesses surfaced new bugs, without giving a count.

**Why it matters:** If an agent writes the proof and a checker accepts it, the specification becomes the part a human still has to review. A verified label is only as broad as the harness it rests on, so anyone putting agent-produced proofs into a verification loop should check which call sequences the theorem actually covers.
