AINEWS
Pesce's column, Anthropic's Mythos multi agent security project, and an open source pipeline converge on the same finding: agents are most useful as hostile quality assurance testers, not as autonomous authors.
For a week straight, an AI agent found a new way to break Mark Pesce's 30-year-old equipment. Every hour, a new failure. The Register columnist's response was not to fire the agent. He reframed the job. The agent was not there to author. It was there to be the most demanding tester on the team.
That reframe is the portable pattern. Most agent pitches in 2026 sell the model as an author, a coder, a junior engineer. The shipped artifacts of the past month sell the opposite: the value of an agent sits in hostile QA, in the loop that does not trust the work and breaks it before a human ships it.
The clearest evidence is from April 7, 2026, when Anthropic announced Project Glasswing and Claude Mythos, a consortium of AI agents built to find and patch vulnerabilities in critical open-source software. Pesce's August 12 column cites the consortium's results: Mythos autonomously rediscovered a 27-year-old OpenBSD bug and a 16-year-old FFmpeg bug, then produced a working FreeBSD NFS remote code execution. The point was not that the agent was clever. The point was that a piece of code no human had read carefully in a decade finally met a tester that never got bored.
The pattern has already shipped as a tool. mythos-agent is an open-source AI code-review assistant with a four-stage pipeline: Recon, Hypothesis, Analyze, Exploit. Each stage is a deliberately hostile pass at the previous stage's output. The pipeline assumes the code is wrong and tries to prove it. It comes with EU Cyber Resilience Act compliance documentation, which is a hint about where the demand is heading. Regulators want the same hostile-QA loop on every maintained project, and most teams cannot staff it.
The byline carries weight, even if the column body is opinion rather than reporting. Pesce co-invented VRML, the Virtual Reality Modeling Language, in 1994, founded one of the first consumer VR startups in 1991, and has written for The Register since 2014. When a futurist who has been wrong about the future in public for thirty years says the agent's job is testing, his track record on the beat supports the claim even when the column itself is not the load-bearing reference.
The standard objection is that agents hallucinate. They invent APIs that do not exist, return plausible but wrong answers, and confidently explain code that does not run. In the author role, hallucination is a defect that ships to production. In the tester role, hallucination is just another break case. The agent that invents a fake endpoint during hostile QA has produced a new test, not a new feature. The bug surface grows. The shipped surface stays clean.
A benchmark at aisle.com ran Mythos's flagship FreeBSD exploit against eight small open-weights models. All eight recovered the chain, including a 3.6-billion-active-parameter model running at $0.11 per million tokens and a 5.1-billion-active open model that recovered the OpenBSD chain as well. The moat is the system, not the model. A small, cheap agent in a hostile-QA loop matched a frontier system on the test that mattered.
The audit pattern for a team on Monday morning is short. Pick the loop on the project that currently has the weakest tester: legacy integration, dependency upgrades, schema migrations, third-party API drift. Run an agent in that loop with the explicit prompt: assume the code is wrong, find a way to break it, write a reproducer. Keep a human in the author seat and the shipper seat. The agent does not decide what gets built or what gets released. It decides what does not.
That is the role worth hiring for. It is also the role that does not need a frontier model, a vendor, or a press release. It needs a piece of 30-year-old equipment, an agent with a weekend, and a human who reads the report.