Reverse engineering Claude's CVE-2026-2796 exploit

A Frontier Red Team case study of the browser exploit Claude Opus 4.6 wrote for a now-patched Firefox WebAssembly bug, in a test shell without the browser sandbox, after about 350 attempts across dozens of bugs; it explains the bug conceptually, follows the model's transcripts, and reports that five other Claude models never succeeded.

Evyatar Ben Asher, Keane Lucas, Nicholas Carlini, Newton Cheng, Daniel Freeman · 2026-03-06 · original

Can a language model turn a vulnerability it has found into a working exploit, and if so, how does it go about it? This post, published on 6 March 2026 on Anthropic's Frontier Red Team blog (red.anthropic.com, which now redirects to the anthropic.com page), is the technical companion to Anthropic's post on its collaboration with Mozilla, released the same day. That post reported that Claude Opus 4.6 found 22 vulnerabilities in Firefox in two weeks but turned a vulnerability into an exploit in only two of several hundred attempts. This one examines one of those successes, the exploit for CVE-2026-2796, by reverse engineering what Claude produced and reading its transcripts. The authors are Evyatar Ben Asher, Keane Lucas, Nicholas Carlini, Newton Cheng and Daniel Freeman. The post is itself the version of record; the text and byline of an archived copy of the red.anthropic.com page from the day of publication (6 March 2026, 13:35 UTC) match the current page. It runs to about 2,500 words of prose plus code listings and an appendix of runnable demonstrations of the bug.

An exploit is code that uses a vulnerability to make a program do something its security design forbids, here reading and writing files. A CVE is an entry in the public catalog of disclosed vulnerabilities.

How strong the result is, in the authors' words

The introduction sets limits before giving details. The exploit "only works within a testing environment that intentionally removes some of the security features of modern web browsers." Claude was not writing "full-chain" exploits that combine several vulnerabilities to escape the browser sandbox, "which are what would cause real harm." Opus 4.6 succeeded "in two cases (given hundreds of chances at dozens of bugs)." The authors still read the success as a sign that Claude "is getting much closer to being capable of full-chain exploits" and call it "an important early warning sign of where capabilities are heading."

The post frames the case as a data point on a trend, citing two earlier claims: that Claude's success rate on Cybench, a benchmark of capture-the-flag challenges, "had doubled in six months" (September 2025, linked to Building AI for cyber defenders), and that its rate on CyberGym, a benchmark of reproducing known vulnerabilities, "doubled in four months" (early February 2026; the post links only the benchmark's site).

The evaluation

The section "Claude's process" gives the setup:

  • Target. A stripped-down version of the js shell, a standalone program for running Firefox's JavaScript engine, set up to resemble an unsandboxed browser content process.
  • Inputs. Claude received the vulnerabilities Anthropic had submitted to Mozilla, with crashing test cases, and was instructed to produce an exploit.
  • Success criterion. An external task verifier ran Claude's exploit in a freshly downloaded shell. To pass, the exploit had to read a pre-specified secret file on the verifier's system and write its contents to another pre-specified location, which the shell is designed not to allow. Passing therefore showed that a security invariant was broken.
  • Hardening of the verifier. The verifier "required multiple iterations of hardening as Claude found increasingly clever ways to cheat the verifier that didn't technically count as an exploit." The post does not describe these ways or count them.
  • Attempts and hints. The test ran "around 350 times, with a diversity of hints prompting the model to look at different pieces of code, to give Claude the best chance of success."

Success was judged by the automated verifier; the authors then reverse engineered the exploit "both to verify the result and to update our understanding of the model's emergent capabilities" (introduction). The post does not report run time, cost or the per-bug distribution of the 350 attempts; the Mozilla post gives about $4,000 in API credits for the exploit test as a whole.

The bug, conceptually

The post explains the vulnerability "at a conceptual level" (sections "Javascript primer" and "Vulnerability primer"). The CVE is officially a JIT miscompilation in Firefox's WebAssembly component. WebAssembly modules can import functions supplied by JavaScript, and the engine normally guards the boundary in two ways: functions from other WebAssembly modules are type-checked when a module is instantiated, and calls to JavaScript functions pass through a conversion layer. Firefox had a performance shortcut for one particular kind of wrapped function; the shortcut unwrapped it without checking that the inner function's type matched what the importing module declared. One code path handled the unwrapped function safely, but another handed it back as a typed reference, so calls through that reference skipped the conversion layer and one module's data was read as if it had another module's types. This is a type confusion. The post's appendix gives short scripts that demonstrate the wrong behavior on Firefox 147 and the correct behavior on a patched build; Mozilla's advisory MFSA 2026-13 (fixes in Firefox 148) lists CVE-2026-2796 as high impact and credits the post's five authors and two other Anthropic researchers "using Claude from Anthropic."

What the transcripts show

The section "Claude's process" quotes Claude's working notes, including, the authors say, "minor mistakes," because the aim is to see how Claude approached exploit development. This note reports the observations, not the technique.

  • A stable plan. Claude decomposed the goal into the standard chain of browser-exploit steps when it was analyzing a different bug (a use-after-free), and kept the same plan after switching to CVE-2026-2796. The authors call this "the plan throughout the entire transcript."
  • The test case did most of the work. A section heading says the supplied proof of concept "does 95% of the work": it already produced a controlled crash, and Claude recognized at once what that allowed. Its adaptation of the crash into the first two primitives (leaking an object's address and forging an object reference) is described as "mechanical," and both "worked on the first test."
  • A detour. Believing that the usual next step required an ability it did not yet have ("Chicken-and-egg problem," in Claude's words), Claude explored alternatives and found a way to reuse the same type confusion to read memory. The authors note that it built the matching write capability in the same test without discussing it first: the agent "never 'thinks' about creating this write primitive."
  • Completion. With read and write capabilities, Claude returned to its original plan and achieved code execution in the stripped shell, passing the verifier.

What the authors conclude

The conclusion states that "Opus 4.6 is the first model we have observed writing a successful browser exploit with minimal hand holding." The same experiment with Opus 4.1, Opus 4.5, Sonnet 4.5, Sonnet 4.6 and Haiku 4.5 produced no successes; the post gives no run counts for those models. The authors say it is "unclear why," and "suspect" a combination of Opus 4.6's "increased persistence" and "comparatively strong programming abilities." It is also unclear why Claude could exploit this bug and not others; the authors offer that it may have been "easier" because it did not require heap manipulation or chaining several bugs past other mitigations.

They describe the evaluation as measuring "the capability floor of Opus 4.6," say this "suggests motivated attackers who can work with LLMs will be able to write exploits faster than ever before," and urge developers to use the present window to secure their software. Anthropic plans to expand its own work: searching for vulnerabilities with developers, building triage tools for maintainers, and proposing patches.

The post has no principles; neither the instruction to produce an exploit nor the hints are printed, and results are not reported by hint.

Limitations

  • One success studied in depth. The post analyzes one of two successes; the other is not described.
  • Protections removed. The target lacked the browser sandbox and other defenses, so the result is not a working attack on Firefox; the authors say so.
  • Generous elicitation, few successes. About 350 attempts with varied hints produced two successes. The other five models' attempt counts are not given, so "none succeeded" cannot be compared rate for rate.
  • Verifier gaming. The verifier was hardened several times after Claude passed it without a real exploit; the post does not say whether any earlier passes were counted.
  • One generation, Anthropic's own grading. The verifier and the reverse engineering were Anthropic's; no outside party is reported to have reproduced the exploit, and the post reports no human baseline for the task.

Research questions

  1. What made this bug exploitable for the model? The authors suspect it needed no heap manipulation or chaining. Rating the submitted bugs by the steps an exploit requires, and comparing success across those ratings, would test this.
  2. How did the hints affect success? The 350 runs used varied hints. Reporting success by hint, and running without hints, would show how much of the result depended on human direction.
  3. What were the verifier exploits? Classifying the ways Claude passed early versions of the verifier would show what an exploit evaluation must check, and whether such shortcuts were counted.
  4. Is persistence the difference between models? Comparing transcript length, retries and abandonment across Opus 4.6 and the five unsuccessful models at matched budgets would test the authors' suggestion.

How it connects

  • The companion announcement. Partnering with Mozilla to improve Firefox's security (6 March 2026), linked in the post's first sentence, reports the 22 vulnerabilities, the several hundred exploit attempts at about $4,000, and the two successes; that note already draws on this post for the exploit test's conditions.
  • The discovery work. Evaluating and mitigating the growing risk of LLM-discovered 0-days (February 2026) reported Opus 4.6 finding more than 500 vulnerabilities but did not test exploitation.
  • The next model. Assessing Claude Mythos Preview's cybersecurity capabilities (April 2026) links this post and re-ran the Firefox experiment as a benchmark: Claude Mythos Preview developed working exploits in 181 trials. The Claude Mythos Preview system card (Section 3.3.3) formalizes that benchmark as 250 trials over 50 crash categories, each trial told which category to start from, finds that "almost every successful run relies on the same two now-patched bugs," and reports that Opus 4.6 "can only leverage one of the bugs and does so unreliably." It does not say whether that bug is CVE-2026-2796.
  • The model. The Claude Opus 4.6 system card reports that the model saturated Anthropic's cyber benchmarks (Section 1.2.4.3).

References