Vulnerability Reports Are Not Special Anymore

What Valsorda Argued

Valsorda's post, published on his personal site at words.filippo.io on June 23, is a short essay in the genre of "an old practitioner admits the rules have changed" ^[1].

His core claim: the workflow security researchers have used for two decades is obsolete. For years, Valsorda wrote, he told new team members that vulnerability reports were "special" and security researchers were "doing us a favor by reporting things confidentially instead of doing full disclosure, so we owe them something." That premise is gone in 2026.

The argument has three parts:

1. Insight is no longer scarce. "LLMs are as good as almost any security researcher, and anyone can run them. The maintainers can run them. The attackers can run them. The insight is not scarce and precious anymore."
2. Confidentiality buys less than it used to. "The attackers don't need to read the full disclosure post to learn about the vulnerability: they can ask their own LLM and, in fact, they also probably have the same triage bottleneck as the defenders do."
3. The bottleneck moved. "The bottleneck now is not finding potential issues but assessing which ones are real. Unless there's already a trust relationship, external researchers can't meaningfully contribute to that triage process."

The conclusion Valsorda draws: "Triage, rapid remediation, and, as ever, prevention are the job now. And we should all figure out how to run LLM analysis in CI, I suppose."

Why It Matters for Surveillance Coverage

This is a security-labor post on a privacy site, so the case for covering it here requires a beat read. The read is direct: the same capability shift that Valsorda is documenting in open-source security work is the same capability shift that the Fable 5 / Mythos 5 family of coverage has been tracking for two months ^[2][3].

The Will It Mythos benchmark, published June 23, 2026, found that Mythos, the latest Anthropic model, can locate vulnerabilities in real code at a level that maps onto the work human security researchers used to own. Mythos did not invent the vulnerabilities. It found them, and the question of what to do with what it found is the same triage bottleneck Valsorda names.

The pattern Valsorda describes from the maintainer side, where external reporters' contributions are now redundant because internal LLM analysis does the same work, is the same pattern from the vendor side. Anthropic's models are not just a research curiosity. They are on a trajectory to replace the labor pool that the open-source and security-industrial complex has relied on for twenty years. The Mythos conversation is the AI-vendor version. Valsorda's post is the maintainer version.

How the Hacker News Thread Read It

Valsorda's post hit the front page of Hacker News within hours of publication and stayed there. The thread crossed 350 points at 17 hours old, with a sustained comment velocity of 12 comments per hour, the highest sustained comment-velocity of any security-research-labor thread on HN in the tracker record ^[4].

The split in the comments is informative. The dominant frame, especially from senior security engineers, is "he's right and the implication is uncomfortable." The minority frame, mostly from vulnerability disclosure platform staff and bug bounty operators, is "the LLM plateau is temporary and the profession will re-stabilize at a new higher bar." Valsorda himself added nuance in a Bluesky exchange, replying to Avery Pennarun: "There's been a step change in ability to find vulns, but the only stable outcome (once we get there) is fewer vulns getting released. When that happens there will be a new higher bar and finding them will be hard again."

That is the most interesting sentence in the whole exchange. Valsorda is not claiming LLM vulnerability discovery will permanently devalue the profession. He is claiming the current dynamic, where LLMs find vulns faster than researchers can triage them, will persist as long as the models keep getting better. The plateau question, whether the models stop improving at this task, is the empirical question that will determine whether his prediction holds.

Who Filippo Valsorda Is

The author matters here because the claim is structural, not promotional. Filippo Valsorda led the Go Security team at Google, was the cryptography lead at Cloudflare, and is a long-time contributor to the Go standard library's crypto packages. He is not a vendor employee selling a product, not a researcher publishing a paper that needs citations, and not a journalist with a frame to push. He is a practitioner with twenty years of security work behind him telling other practitioners that the deal they have been operating under is broken ^[1].

That makes the post authoritative in a way that vendor AI safety claims rarely are. Valsorda is not arguing for or against any specific model. He is describing what he sees as the working conditions of security research in 2026 and predicting where they are going.

The Labor Angle Is the Surveillance Angle

The closest parallel on this site is the coverage of the Model Capability Initiative at Meta, the program that recorded every keystroke from US employees' work laptops to train AI agents that automate office work ^[5]. The same AI capability shift that Meta's CTO Andrew Bosworth cited when justifying MCI is the same capability shift Valsorda is documenting from the maintainer side. Bosworth told employees that MCI would help build "agents that primarily do the work" while employees "direct, review and help them improve." Valsorda is telling maintainers that the LLM triage pipeline does the work that external security researchers used to do.

The pattern across both stories is the same: AI does work that a specific labor pool used to own, and the labor pool is told that the new division of labor still has a role for them in "directing, reviewing, and improving" the AI's output. The MCI story is the corporate version. Valsorda's post is the open-source version. Both are the same AI-labor-displacement story from different sides.

What to Watch

• The next model release. Valsorda's prediction rests on the assumption that LLM vulnerability discovery keeps improving. If Mythos 5 or GPT-6 or Gemini 2 ships in the next quarter with materially better vulnerability-finding capability, the structural shift he describes accelerates. If the next release plateaus, his prediction is wrong on a five-year horizon and right on a one-year horizon.
• The bug bounty platforms. HackerOne, Bugcrowd, and Intigriti all run businesses that intermediate between reporters and maintainers. If external reports stop being scarce, the platforms either have to change what they sell or accept that the value they capture is shrinking. Watch for platform commentary on the post.
• The CI integration question. Valsorda's prescription, "we should all figure out how to run LLM analysis in CI," is a product question. The vendor that wins the next eighteen months is whichever one gets LLM-based vulnerability analysis integrated into standard CI pipelines first.
• The CVE backlog. If LLM-generated vulnerability findings flood the disclosure system faster than humans can triage them, the CVE / NVD backlog becomes the visible symptom. Watch for NVD publication delays or CVE Numbering Authority policy changes.
• The Mythos benchmark follow-ups. The Will It Mythos post is the first empirical anchor. The next question is whether Mythos finds the same vulnerabilities independent security researchers do, in blind testing, on code the model has not been trained on. Watch for that follow-up.

The Bottom Line

Filippo Valsorda, one of the most respected security practitioners of his generation, says the work security researchers have done for two decades is now done by LLMs. He is not making a vendor argument. He is making a structural observation about where the labor value sits in 2026.

The traditional deal was: researchers found vulnerabilities, maintainers fixed them quietly, attackers learned about the vulnerabilities after the fix shipped. The new deal is: LLMs find vulnerabilities for everyone, including attackers, the bottleneck is triage, and the value of external disclosure is collapsing into a trust-relationship question rather than an information question.

The surveillance angle is the displacement angle. The same LLM capability shift that Valsorda is documenting on the maintainer side is the shift Meta is operationalizing on the employee side. Both stories point at the same broader pattern: AI capability is moving past the labor pool it was supposed to augment, and the people affected are being told to focus on directing and improving the output. That pitch has not gotten more convincing since April.