Will AI Replace Security Researchers? What Changes Now

AI will not replace security researchers, but it is changing what the job looks like. Models such as Anthropic’s Claude Mythos Preview can now find and chain software vulnerabilities into working exploits with little human input, which shifts researcher time away from manual discovery and toward verification, triage, and the judgment calls a model cannot make on its own.

What Changed: Autonomous Vulnerability Discovery at Machine Speed

Anthropic’s Project Glasswing put the shift in concrete numbers rather than a general claim about “AI getting better.” Where the company’s previous flagship, Opus 4.6, had a near-zero autonomous exploit success rate on internal benchmarks, Mythos Preview reached 72.4%. Anthropic’s own red team documented the model autonomously identifying and exploiting a 17-year-old remote code execution flaw in FreeBSD’s NFS server, alongside decades-old bugs in OpenBSD and FFmpeg that had survived years of human and automated review.

A month into the program, Anthropic reported that roughly 50 partner organizations using the model had collectively found more than 10,000 high- or critical-severity vulnerabilities, with several partners’ bug-finding rate increasing more than tenfold. Mozilla alone found and fixed 271 vulnerabilities in a single Firefox release using the model, more than ten times what it found in the previous release using Opus 4.6.

That scale is the actual story, not the existence of AI-assisted vulnerability discovery, which is not new. Google’s Big Sleep project surfaced a real SQLite vulnerability through AI assistance in 2024. What changed with Mythos-class models is that discovery, exploit chaining, and proof-of-concept generation now happen with little to no human involvement in between.

Why More Discoveries Doesn’t Mean Fewer Researchers

Anthropic’s own framing of the Glasswing results makes the researcher-replacement question easier to answer than most commentary suggests: the company has said openly that finding new vulnerabilities is no longer the bottleneck it used to be, and that verifying, disclosing, and patching what gets found is now the limiting factor. Finding a vulnerability faster does not shrink the work of confirming it is real, understanding its business impact, and deciding what to do about it. It moves that work to the front of the queue and makes there more of it.

The Glasswing numbers show this directly. Of roughly 6,200 vulnerability candidates the model flagged across scanned open-source projects, human expert review confirmed 1,726 as genuine, exploitable flaws, and 1,094 of those as high or critical severity. Most of a model’s raw output still needs a person to separate the real finding from the plausible-looking false positive, at a volume no security team was staffed for a year ago.

Separately, Wordfence reported that AI-assisted vulnerability submissions to its bug bounty program grew 453% between October 2025 and April 2026, forcing the company to build a multi-agent triage pipeline just to keep pace with intake, on top of the more than fourfold growth in report volume the program had already seen. The bottleneck has not disappeared. It has relocated from discovery to verification, and verification still requires a person who understands what the finding actually means in context.

The Three Places Humans Still Decide

Three parts of the security lifecycle do not move to the model layer no matter how capable that model becomes.

Initial access hasn’t changed. A zero-day exploit chain still needs a way into a target system, and phishing, business email compromise, and social engineering remain the dominant way attackers get one. A Mythos-class model can hand an attacker a working exploit; it still generally needs a human somewhere to click, approve, or trust something they should not, unless the exploit targets an internet-facing service directly. Faster vulnerability discovery gives attackers a better payload. It does not remove the delivery problem defenders have been managing for two decades.

Verification and prioritization stay human. A finding is not a fixed vulnerability. Someone has to reproduce it, assess what it actually exposes, and decide where it sits in a patch queue against everything else competing for the same engineering time. The NIST incident response framework already treats triage as a distinct phase for exactly this reason: detection volume and response capacity are different constraints, and AI has only widened the gap between them.

Accountability stays with a named person. When an organization deploys an AI system, agentic or otherwise, inside its own environment, someone has to own what that system produces and be answerable for it. That ownership question does not resolve at the model layer in any current AI system, Mythos-class or otherwise.

What This Means for Vulnerability Management Right Now

A zero-day exploit’s lifecycle has always run on two independent clocks: how fast an attacker can weaponize a flaw, and how fast a defender can patch it. Mythos-class models compress the first clock dramatically. The second clock, gated by testing, change management, and engineering bandwidth, has not moved at the same pace, and that gap is where organizational risk is actually growing.

Three adjustments follow directly from that gap:

• Patch velocity becomes a board-level metric, not just an operations one. A defender who patches on a monthly cycle is defending against last year’s attacker capability, not this year’s.
• Triage capacity needs headcount and tooling, not just detection tooling. A vulnerability scanner that finds ten times more candidates is only useful if verification capacity grows with it.
• AI governance inside the security team is now in scope, not just AI governance for the rest of the business. A security team running its own AI-assisted scanning tools has the same shadow AI and oversight questions as any other department, and insider threat programs increasingly need to account for AI tool use by the people defending the network, not only the people being defended.

What Security Researchers’ Jobs Actually Look Like Now

Practitioner surveys back up what the Glasswing numbers imply. HackerOne’s ninth Hacker-Powered Security Report found that agentic AI is improving baseline coverage without replacing expert insight, with architecture and business-logic flaws, the category AI is worst at finding, growing in relative importance as the easier bugs get automated away. Researchers who adapt are shifting toward orchestrating AI tools, validating their output, and specializing in the logic and context flaws that require understanding how a specific business actually operates, not just how its code is structured.

That shift mirrors what happened to other technical fields when automation took over the mechanical parts of the job: the remaining human work concentrates in judgment, context, and accountability, and it becomes more valuable per hour even as the total headcount doing purely mechanical discovery shrinks. It is the same pattern already showing up on the development side, where AI coding assistants shift risk toward the judgment calls a model cannot reliably make on its own.

Building a Human-Centered Response to AI-Accelerated Threats

None of this reduces the importance of the human layer. Security programs that measure human risk rather than treating awareness as a compliance checkbox are better positioned for exactly this shift, because the two things that do not change with model capability are the initial-access vector and the accountability chain, and both of those run through people. Human error remains the largest single factor in most breaches regardless of how the technical layer evolves, which is exactly why that layer needs more investment now, not less.

Two concrete moves matter most:

• Keep testing the initial-access vector directly. Since phishing and social engineering remain the entry point regardless of how good AI gets at finding zero-days, Threatcop’s TSAT baselines exactly how exposed your workforce is to that entry point, through simulation and a per-employee vulnerability score, rather than assuming awareness training is working because nobody has complained.
• Build SOC processes that assume higher finding volume. A triage workflow designed for last year’s report rate will buckle under this year’s, regardless of how good the underlying detection has become.

The Bottleneck Moved. It Didn’t Disappear

Mythos-class models change how fast a vulnerability gets found. They do not change who decides whether it matters, who fixes it, or who a phishing email still has to fool to get an attacker in the door in the first place. Organizations that treat this as a reason to invest less in their people are reading the data backwards: the technical layer just got faster, which makes the human layer the part of the defense that is no longer keeping pace by default.

If your security program needs to catch up on the AI side of this shift as much as the human side, see how AI is affecting cybersecurity more broadly, including the agent-specific risks that sit alongside the vulnerability-discovery risks covered here.

Frequently Asked Questions

Adhish Chakma

Adhish Chakma is a Senior Product Manager at Kratikal, where he leads product initiatives focused on cybersecurity and AI-powered solutions. With experience in product management and cybersecurity, he works on developing practical technologies that address evolving security challenges. His areas of interest include People Security Management, cybersecurity awareness, AI-driven security, email security, and human-layer risk. He is passionate about building security products that make organizations more resilient against emerging cyber threats.

Adhish Chakma

Adhish Chakma is a Senior Product Manager at Kratikal, where he leads product initiatives focused on cybersecurity and AI-powered solutions. With experience in product management and cybersecurity, he works on developing practical technologies that address evolving security challenges. His areas of interest include People Security Management, cybersecurity awareness, AI-driven security, email security, and human-layer risk. He is passionate about building security products that make organizations more resilient against emerging cyber threats.