model releaseAnthropic

Anthropic's Mythos model finds thousands of high-severity bugs in Firefox, including 15-year-old vulnerabilities

TL;DR

Mozilla's Firefox team reports that Anthropic's Mythos model has discovered thousands of high-severity security vulnerabilities, including bugs that had remained undetected for more than 15 years. In April 2026, Firefox shipped 423 bug fixes compared to just 31 in April 2025, marking a 13x increase attributed to AI-assisted vulnerability detection.

3 min read
0

Anthropic's Mythos Finds Thousands of Firefox Vulnerabilities

Mozilla's Firefox security team has disclosed that Anthropic's Mythos model discovered thousands of high-severity bugs in the browser's codebase, including vulnerabilities that had existed undetected for over 15 years.

The results represent a dramatic shift in AI-assisted security research. In April 2026, Firefox shipped 423 bug fixes, compared to 31 in April 2025—a 13-fold increase. Mozilla attributes this acceleration directly to Mythos' capabilities.

Technical Capabilities

According to Mozilla researchers, Mythos demonstrated particular strength in finding sandbox vulnerabilities—among the most difficult security flaws to detect. These require multi-step exploits: the model must write a compromised browser patch, then use that code to attack Firefox's sandbox isolation system.

Mozilla's bug bounty program pays up to $20,000 for sandbox vulnerabilities, its highest reward tier. Brian Grinstead, a distinguished engineer at Mozilla, told TechCrunch that Mythos is finding more sandbox issues than human researchers ever did, despite the substantial financial incentive.

The 12 published bugs include:

  • Multiple sandbox escape vulnerabilities
  • A 15-year-old HTML parsing error
  • Various high-severity security flaws across the codebase

Agentic Filtering Reduces False Positives

Mozilla credits two factors for the breakthrough: improved model capabilities and agentic systems that self-assess and filter out false positives. Previous AI security tools generated too many low-quality reports to be practical.

"It is difficult to overstate how much this dynamic changed for us over a few short months," Mozilla researchers wrote. "First, the models got a lot more capable. Second, we dramatically improved our techniques for harnessing these models."

Human Review Still Required

Despite Mythos' detection capabilities, Mozilla is not using AI to fix the bugs. While the model generates proposed patches, Firefox engineers write and review all deployed code manually. "For the bugs we're talking about in this post, every single one is one engineer writing a patch and one engineer reviewing it," Grinstead said.

Disclosure and Impact

Anthropic previewed Mythos in April 2026 but withheld public release until critical vulnerabilities could be patched. The company has followed responsible disclosure protocols, though security researchers note that adversaries may be using similar techniques with less capable models.

Anthropic CEO Dario Amodei argued the tools favor defenders: "If we handle this right, we could be in a better position than we started, because we fixed all these bugs. There are only so many bugs to find."

Grinstead offered a more cautious assessment: "It's useful for both attackers and defenders, but having the tool available shifts the advantage a little bit to defense. Realistically, nobody knows the answer to this yet."

What This Means

Mythos represents the first AI system demonstrating practical superiority over human researchers at finding complex security vulnerabilities at scale. The 13x increase in Firefox's monthly bug fixes suggests we're entering a period where legacy codebases will undergo unprecedented security scrutiny. The critical unknown: whether attackers or defenders benefit more as these capabilities proliferate. Mozilla's experience suggests that even with access to state-of-the-art detection, human expertise remains essential for remediation.

Related Articles

research

Anthropic mathematician Levent Alpöge claims Claude-assisted proof of the Riemann hypothesis, posts it in Lean

Levent Alpöge, a mathematician employed by Anthropic, says he has proven the Riemann hypothesis, first proposed in 1859, working with Claude. He posted the proof coded in Lean, and experts have yet to confirm that it is correct.

product update

Anthropic Adds Experimental Code Mode to Claude Code v2.1.323, Chaining Tool Calls in JavaScript

Anthropic has added an experimental Code Mode to Claude Code v2.1.323. Claude can write JavaScript that chains Bash, Read and Edit calls in one step, and keep talking while long commands run.

model release

Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context

Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.

model release

NVIDIA releases Nemotron-Labs-Sigma: 14 open 3B and 8B continuous diffusion language model checkpoints

NVIDIA released Nemotron-Labs-Sigma, a set of 14 open continuous diffusion language model checkpoints at 3B and 8B parameters. According to MarkTechPost, Sigma denoises text in a learned 16-D embedding space, distills to few-step samplers, and decodes up to 4.20x faster than autoregressive decoding.

Comments

Loading...