· Mar 6, 2026
Anthropic reports Claude finding Firefox vulnerabilities with Mozilla
Markdown · JSON · Original source
published this account on March 6, 2026, describing a collaboration with Mozilla in which Opus 4.6 found vulnerabilities in Firefox. Anthropic says Mozilla classified some findings as high severity and fixed most reported issues in Firefox 148. The account describes Mozilla helping Anthropic decide what to report, including submitting findings in bulk without validating each one; Anthropic says the collaboration helped it adjust its approach. Anthropic also reports that Claude could turn vulnerabilities into working exploits in only a small number of its tests. Those exploits worked only in a test environment that lacked some security features found in modern browsers, including the sandbox. Anthropic says this leaves vulnerability discovery ahead of exploitation for now, but warns that the gap may not last. Its conclusion is that developers should use this window to improve software security, while Anthropic plans to expand its work with developers on finding, triaging and fixing vulnerabilities. How quickly, if at all, might models’ ability to exploit vulnerabilities catch up with their ability to find them?
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Mar 6, 2026 (2026-03-06T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-07T23:08:23.641Z. First observed: 2026-10-19T02:33:19.503Z. Retrieved: 2026-10-28T23:05:40.389Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 28, 2026; originally published March 6, 2026. The source page is marked as edited after publication, but the retrieved evidence does not establish which passages, if any, were added later.
Open question
How quickly, if at all, might models’ ability to exploit vulnerabilities catch up with their ability to find them?
Source-stated
Anthropic says Claude Opus 4.6 found 22 Firefox vulnerabilities during a two-week collaboration with Mozilla, which assigned 14 a high severity rating.
Claude Opus 4.6 discovered 22 vulnerabilities over the course of two weeks. Of these, Mozilla assigned 14 as high-severity vulnerabilities
Source-stated
Anthropic says Mozilla fixed most of the issues in Firefox 148, with the rest planned for upcoming releases.
Most issues have been fixed in Firefox 148, with the remainder to be fixed in upcoming releases.
Source-stated
Anthropic reports that Opus 4.6 successfully turned vulnerabilities into exploits in two cases during tests run several hundred times.
We ran this test several hundred times with different starting points, spending approximately $4,000 in API credits. Despite this, Opus 4.6 was only able to actually turn the vulnerability into an exploit in two cases.
Source-stated
Anthropic says the exploits worked only in its testing environment, which lacked some security features found in modern browsers, including the sandbox.
The exploits Claude wrote only worked on our testing environment, which intentionally removed some of the security features found in modern browsers. This includes, most importantly, the sandbox
Source-stated
Anthropic warns that the gap between vulnerability discovery and exploitation may not last, and urges developers to use the present window to improve software security.
it is unlikely that the gap between frontier models’ vulnerability discovery and exploitation abilities will last very long. If and when future language models break through this exploitation barrier
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 5e1db1c6-4016-5162-17e5-0c0826c77512.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T23:03:48.199Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 2e04e046532f90e7724cec6615387bcdf3fa1dbe33ff1eabb782a0577a13174e. Source text hash: 4ed371864ba1802205963f372569ce20c958ee21f5e27055822390894ccc4d69.
Revision and source-check history
Version 1 · 2026-11-07T23:08:23.641Z
Anthropic reports Claude finding Firefox vulnerabilities with Mozilla
published this account on March 6, 2026, describing a collaboration with Mozilla in which Opus 4.6 found vulnerabilities in Firefox. Anthropic says Mozilla classified some findings as high severity and fixed most reported issues in Firefox 148. The account describes Mozilla helping Anthropic decide what to report, including submitting findings in bulk without validating each one; Anthropic says the collaboration helped it adjust its approach. Anthropic also reports that Claude could turn vulnerabilities into working exploits in only a small number of its tests. Those exploits worked only in a test environment that lacked some security features found in modern browsers, including the sandbox. Anthropic says this leaves vulnerability discovery ahead of exploitation for now, but warns that the gap may not last. Its conclusion is that developers should use this window to improve software security, while Anthropic plans to expand its work with developers on finding, triaging and fixing vulnerabilities. How quickly, if at all, might models’ ability to exploit vulnerabilities catch up with their ability to find them?
Original source
· Article
Source details
Original source date: Mar 6, 2026 (2026-03-06T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-07T23:08:23.641Z. First observed: 2026-10-19T02:33:19.503Z. Retrieved: 2026-10-28T23:05:40.389Z.
Limits
Text as retrieved on October 28, 2026; originally published March 6, 2026. The source page is marked as edited after publication, but the retrieved evidence does not establish which passages, if any, were added later.
Open question
How quickly, if at all, might models’ ability to exploit vulnerabilities catch up with their ability to find them?
Source-stated
Anthropic says Claude Opus 4.6 found 22 Firefox vulnerabilities during a two-week collaboration with Mozilla, which assigned 14 a high severity rating.
Claude Opus 4.6 discovered 22 vulnerabilities over the course of two weeks. Of these, Mozilla assigned 14 as high-severity vulnerabilities
Source-stated
Anthropic says Mozilla fixed most of the issues in Firefox 148, with the rest planned for upcoming releases.
Most issues have been fixed in Firefox 148, with the remainder to be fixed in upcoming releases.
Source-stated
Anthropic reports that Opus 4.6 successfully turned vulnerabilities into exploits in two cases during tests run several hundred times.
We ran this test several hundred times with different starting points, spending approximately $4,000 in API credits. Despite this, Opus 4.6 was only able to actually turn the vulnerability into an exploit in two cases.
Source-stated
Anthropic says the exploits worked only in its testing environment, which lacked some security features found in modern browsers, including the sandbox.
The exploits Claude wrote only worked on our testing environment, which intentionally removed some of the security features found in modern browsers. This includes, most importantly, the sandbox
Source-stated
Anthropic warns that the gap between vulnerability discovery and exploitation may not last, and urges developers to use the present window to improve software security.
it is unlikely that the gap between frontier models’ vulnerability discovery and exploitation abilities will last very long. If and when future language models break through this exploitation barrier
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 5e1db1c6-4016-5162-17e5-0c0826c77512.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T23:03:48.199Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 2e04e046532f90e7724cec6615387bcdf3fa1dbe33ff1eabb782a0577a13174e. Source text hash: 4ed371864ba1802205963f372569ce20c958ee21f5e27055822390894ccc4d69.
uncertain · 2026-11-07T23:03:48.199Z · anthropic/claude-opus-5.5
Publication: 5e1db1c6-4016-5162-17e5-0c0826c77512
attribution — pass: The reading consistently attributes the account, figures and warnings to Anthropic, which published the post. Mozilla's actions, including its severity ratings, fixes and bulk-submission advice, are presented as Anthropic's description of them. The source is a single article, not a listing.
In this post, we share details of a collaboration with researchers at Mozilla in which Claude Opus 4.6 discovered 22 vulnerabilities
summary — uncertain: The reading faithfully covers the collaboration, the limited exploit results, the caveat about the test environment lacking the sandbox, and Anthropic's warning about a closing window. However, central terms are not explained and the definitions list is empty: exploit, sandbox and triaging. The source glosses some terms itself, such as zero-day and Use After Free. A general reader may not grasp the difference between finding a vulnerability and exploiting it, which the question depends on. The smallest repair is a plain-language gloss or term-library tooltip for exploit and sandbox. The question is substantive and grounded in the source.
we wanted to understand whether Claude could also develop the sorts of tools that a hacker would use to take advantage of these bugs to execute malicious code
date — pass: The stated date of March 6, 2026 matches the UTC calendar day of evidence.originalPublished. The evidence is a publisher page, so no archive disclosure is needed. The forecast that the gap will close is reported as Anthropic's warning, not as an outcome. The page was edited after publication, which the limitations disclose. The cited passages (Firefox 148, February 2026 findings, exploit tests) are consistent with the original March piece. The reading does not use the later-sounding passage saying Mozilla has since begun experimenting internally.
Claude Opus 4.6 found 22 vulnerabilities in February 2026, more than were reported in any single month in 2025.
support — pass: Each claim matches its quoted text: 22 vulnerabilities with 14 rated high severity, most fixes shipped in Firefox 148, exploits produced in two cases over several hundred runs, and exploits that worked only without the sandbox. The warning about the gap and the call for developers to act are also quoted accurately. Anthropic's planned expansion is supported by the source. These cited passages appear to be core parts of the original article rather than later edits. Grounding in the source does not independently verify Anthropic's claims.
We urge developers to take advantage of this window to redouble their efforts to make their software more secure.
media — pass: No media was selected and no verified media candidates were supplied, so there is no identity or relevance issue to check.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore Anthropic