Assessing Claude Mythos Preview’s cybersecurity capabilities

The Frontier Red Team's technical account, released with Claude Mythos Preview and Project Glasswing, of a model that the authors say can find and exploit zero-day vulnerabilities in every major operating system and web browser when directed to, often autonomously; it reports benchmark jumps over Claude Opus 4.6 (working Firefox shell exploits in 181 of 250 trials against under 1%), case studies, a check of the model's severity ratings against human validators, and advice to defenders, with most findings withheld behind cryptographic commitments pending disclosure.

Nicholas Carlini, Newton Cheng, Keane Lucas et al. · 2026-04-07 · original

What can a model do in computer security when it can find and exploit previously unknown vulnerabilities in the most heavily tested software? This post, published on 7 April 2026 on Anthropic's Frontier Red Team blog (red.anthropic.com, which now redirects to the anthropic.com page), is the technical companion to the announcement of Claude Mythos Preview and of Project Glasswing, the program under which Anthropic offered the model to a limited group of industry partners and open-source developers instead of releasing it generally. The authors write that they hope it will show "why we view this as a watershed moment for security." The post is long (about 13,000 words before its appendix of hashes), and most of its findings are withheld: over 99% of the vulnerabilities found had not been patched, so they are described in the abstract and committed to by cryptographic hash.

The byline lists Nicholas Carlini, Newton Cheng, Keane Lucas, Michael Moore, Milad Nasr, Vinay Prabhushankar and Winnie Xiao, followed by 19 further contributors; a note at the end says the author list was updated on 9 April 2026. An archived copy of the red.anthropic.com page from the day of publication (7 April 2026, 18:09 UTC) lists 15 further contributors (David Forsythe, Kyla Guru, Mojtaba Mehrara and Renee Nichol were added) and otherwise has the same text. The post is the version of record for its case studies. The Claude Mythos Preview system card (Section 3), released the same day, is the version of record for the benchmark setups, and gives conditions the post leaves out; both are cited below by location.

Terms used below: a zero-day is a vulnerability unknown to the software's maintainers; an N-day is one that has been disclosed and patched but may remain exploitable on systems not yet updated. An exploit turns a vulnerability into an attacker's goal, such as running code or gaining administrator ("root") access. Privilege escalation gains higher permissions on a machine where the attacker already has some.

The headline claims

The section "The significance of Claude Mythos Preview for cybersecurity" opens with the post's main claims, stated as observations from about a month of testing:

  • Mythos Preview "is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so." Many of the bugs are ten or twenty years old; the oldest found so far was a now-patched 27-year-old bug in OpenBSD.
  • Its exploits include a browser exploit that chained four vulnerabilities to escape both the renderer and operating-system sandboxes, local privilege escalation on Linux and other systems, and a remote code execution exploit for FreeBSD's NFS server granting root to unauthenticated users.
  • People without security training could use it: engineers at Anthropic "with no formal security training" asked for remote code execution vulnerabilities overnight and found "a complete, working exploit" in the morning. This is reported as an anecdote, without counts.
  • The capabilities were not trained for explicitly. The authors write that they "emerged as a downstream consequence of general improvements in code, reasoning, and autonomy," and that the improvements that make the model better at patching also make it better at exploiting.

The post argues that, as with fuzzers, language models will in the long run "benefit defenders more than attackers," but that the transition "may be tumultuous," its stated reason for releasing the model first to defenders through Project Glasswing.

Benchmarks: the jump from Claude Opus 4.6

Firefox shell exploitation. The post recalls the March 2026 result that Claude Opus 4.6 turned the Firefox vulnerabilities it found into exploits "only two times out of several hundred attempts" (see Reverse engineering Claude's CVE-2026-2796 exploit). Re-run as a benchmark, Mythos Preview "developed working exploits 181 times, and achieved register control on 29 more" (introduction). The post's chart prints percentages: working exploits in 72.4% of trials and register control (partial control without a full exploit) in a further 11.6% for Mythos Preview, against register control in 14.4% for Opus 4.6 and 4.4% for Claude Sonnet 4.6. Its caption says that the Opus 4.6 exploit rate plotted, "less than 1%," is the one "noted" in "a previous blog," so that bar comes from the earlier, hint-assisted test rather than the same benchmark run, and that Mythos Preview succeeds "nearly 100 times more often." The system card supplies the setup (Section 3.3.3): 50 crash categories found by Opus 4.6 in Firefox 147, five trials per category, 250 trials in all (by this note's arithmetic, 181 and 29 are 72.4% and 11.6% of 250); the target is a shell mimicking a Firefox 147 content process "without the browser's process sandbox and other defense-in-depth mitigations" (also the post's footnote 1); grading is automated at three levels (no progress, controlled crash, code execution). The card adds a condition the post omits: "almost every successful run relies on the same two now-patched bugs." With those two bugs removed, Mythos Preview still outperformed the earlier models, and Sonnet 4.6 did better than in the default version (Section 3.3.3, Figure 3.3.3.B). The card sums up that Mythos Preview leveraged "four distinct bugs" to reach code execution, against Opus 4.6, "which can only leverage one of the bugs and does so unreliably."

OSS-Fuzz crash ladder. Anthropic runs its models against roughly a thousand open-source repositories from the OSS-Fuzz corpus and grades the worst crash each run produces on a five-tier ladder, from a basic crash (tier 1) to complete control-flow hijack (tier 5). With one run on each of roughly 7,000 entry points, Sonnet 4.6 and Opus 4.6 each reached tier 1 in 150 to 175 cases and tier 2 about 100 times, with a single tier-3 crash each; Mythos Preview produced 595 crashes at tiers 1 and 2, "a handful" at tiers 3 and 4, and full control-flow hijack on ten "separate, fully patched targets" (introduction). The post does not say who or what assigned tiers.

Saturated benchmarks. The post says Mythos Preview "mostly saturates" Anthropic's existing benchmarks, which is why the evaluation turned to real zero-days (section "Evaluating Claude Mythos Preview's ability to find zero-days"). The system card gives the numbers: 100% on a 35-challenge Cybench subset with ten trials per challenge (Section 3.3.1), and 0.83 on CyberGym's 1,507 tasks against 0.67 for Opus 4.6 and 0.65 for Sonnet 4.6 (Section 3.3.2).

Why zero-days, and how they were found

The authors argue that known-vulnerability benchmarks cannot separate new capability from memorized solutions, while a zero-day "must be genuine" (footnote 2 reports that in a few N-day cases the model did reference published exploit walkthroughs). The work focused on memory-safety bugs, partly because tools like Address Sanitizer, the authors say, "perfectly separate real bugs from hallucinations."

The scaffold (section "Our scaffold") is the one from Anthropic's earlier vulnerability-finding work: Claude Code with Mythos Preview runs in a container isolated from the Internet, prompted with a paragraph that "essentially amounts to 'Please find a security vulnerability in this program'," and reports either no bug or a bug with a proof of concept. Each agent is pointed at a different file, in an order set by the model's own 1-to-5 ranking of how likely each file is to hold interesting bugs, and a final Mythos Preview agent is asked whether each report is "real and interesting."

The case studies that can be named

Three zero-days are described in detail; the post says each, like "almost all" its findings, was found "without any human intervention after an initial prompt" (section "Finding zero-day vulnerabilities"). This note gives them at the level of what was found and at what cost.

  • OpenBSD, 27 years old. A flaw in the TCP selective-acknowledgment code, added in 1998, combined with an integer-overflow edge case, lets a remote attacker crash any OpenBSD host that responds over TCP. It was the most critical OpenBSD finding of about a thousand scaffold runs, which cost under $20,000 in total and produced several dozen further findings. The run that found it cost under $50, a figure the authors say "only makes sense with full hindsight."
  • FFmpeg, 16 years old. A mismatch between a counter and a sentinel value in the H.264 decoder lets a crafted video write a few bytes out of bounds. The authors judge it not critical and hard to exploit, and note that the underlying pattern dates from 2003 and became a vulnerability in a 2010 refactor, missed since then "by every fuzzer and human who has reviewed the code." Several hundred runs over FFmpeg, costing roughly $10,000, found other vulnerabilities as well; three were fixed in FFmpeg 8.1.
  • A memory-safe virtual machine monitor. An out-of-bounds write from a guest into host memory, in code that must use unsafe operations. The project is not named, and Mythos Preview "was not able to produce a functional exploit."

Several thousand more. Anthropic reports "thousands of additional high- and critical-severity vulnerabilities" in disclosure. Professional contractors manually validate every report before it is sent. On 198 manually reviewed reports, the contractors agreed with the model's severity rating exactly in 89% of cases and within one level in 98% (section "And several thousand more"). From this the authors project, conditionally ("If these results hold consistently"), "over a thousand more critical severity vulnerabilities and thousands more high severity vulnerabilities." The post does not say how the 198 were selected.

Exploits

The section "Exploiting zero-day vulnerabilities" says all exploits discussed target fully hardened systems with defenses enabled (the Firefox benchmark above, per footnote 1, does not), and that the ones it can discuss "are the simplest and easiest to exploit." It reports that expert penetration testers said some exploits written in hours would have taken them weeks, without numbers.

  • FreeBSD remote code execution (CVE-2026-4747). A 17-year-old stack buffer overflow in the NFS server's authentication code lets an unauthenticated remote attacker gain root. Mythos Preview found and exploited it "fully autonomously" over several hours, with the discovery scaffold plus one added sentence asking it to write exploits for triage. The post explains why ordinary defenses did not stop it on this code path, and notes that only an attempted exploit revealed that. For comparison, it says an independent vulnerability research company had shown Opus 4.6 exploiting the same bug, "but succeeding required human guidance." The links go to Calif's write-up (listed in Calif's index on 31 March 2026 as "Claude Wrote a Full FreeBSD Remote Kernel RCE with Root Shell") and its log of 43 user prompts, which shows a person starting from FreeBSD's published advisory and steering setup, debugging and the exploit approach; neither file, as read for this note, names the model version.
  • Linux kernel. Mythos Preview found a number of out-of-bounds write bugs, many of them remotely triggerable, but "even after several thousand scans over the repository" was "unable to successfully exploit any of these," which the authors attribute to the kernel's defense in depth. It did write local privilege escalation exploits, with "nearly a dozen examples" chaining two, three or four vulnerabilities. Details are withheld behind hashes.
  • Web browsers. Mythos Preview found and exploited vulnerabilities "in every major web browser"; details are withheld. Here the post records human involvement: the model "fully autonomously" found the read and write primitives and combined them, and the team "then worked with Mythos Preview to increase its severity," in one case into a cross-origin data bypass and in another, chained with a sandbox escape and privilege escalation, into a webpage that gives an attacker kernel write access. This section, unlike the opening summary, says which steps were joint (see "Limitations").

Logic bugs, cryptography and closed-source software

Mostly without verifiable detail, the post reports logic vulnerabilities (gaps between what code does and what its security model requires) in web applications and the Linux kernel, and weaknesses in widely used cryptography libraries (TLS, AES-GCM, SSH; one, a certificate-authentication bypass, made public that morning). The authors say the model "is able to reliably distinguish between the intended behavior of the code and the actual as-implemented behavior," and concede that for logic bugs "we too lose the ability to (near-)perfectly validate" the reports. For closed-source software the model reconstructed plausible source from stripped binaries and searched it; the reported findings (remote denial of service, smartphone firmware, desktop privilege escalation) are all unpatched and were handled offline under the vendors' bug bounty programs.

N-day exploits

Because zero-day exploits cannot be shown, the post demonstrates comparable skill on patched Linux kernel bugs (section "Turning N-day vulnerabilities into exploits"). The model was given 100 CVEs and known memory-corruption bugs filed against the kernel in 2024 and 2025, filtered them to 40 it judged potentially exploitable, and was asked to write a privilege-escalation exploit for each. More than half of these attempts succeeded. The team selected two to document. Both require NET_ADMIN, a non-default permission. One, starting from a fuzzer report of a one-bit out-of-bounds write, cost under $1,000 at API pricing and took half a day; the other chained a one-byte read from freed memory with a second bug, and took under a day and under $2,000. The authors argue that publishing walkthroughs of bugs patched for over a year adds no risk, and add a disclaimer: they are "very confident in the correctness of the exploits" because the binaries yield root, "less so in our understanding of them." On contamination, they say the exploits are as sophisticated as those written for zero-days, "so we don't believe" prior knowledge explains them.

Disclosure and accountability

Anthropic follows its coordinated vulnerability disclosure principles: high-severity bugs go to professional human triagers before maintainers, so as not to flood maintainers, which also means "fewer than 1%" of potential vulnerabilities had been fully patched (section "Our approach to responsible disclosure"). Recognizing that withheld claims are "difficult to verify," the post publishes SHA-3-224 hashes of 15 reports and proofs of concept (the appendix), promising to replace them with the documents once disclosure completes, "no later than 90 plus 45 days" after reporting. Footnote 3 states the limit of this device: a commitment "does not prove anything about the contents of these files—they could be empty," only that the files existed at the time. The post also commits to state publicly in advance any relaxation of its human-review requirement for reports.

Advice to defenders and the authors' outlook

The section "Suggestions for defenders today" says Mythos Preview will not be made generally available and advises using available models now (Opus 4.6 found high- and critical-severity bugs "almost everywhere we looked"), using models for triage, patching and incident response, and shortening patch cycles, since N-day exploits can now be written "fully autonomously, starting from just a CVE identifier and a git commit hash." The Linux section proposes reexamining defenses whose value comes from friction rather than hard barriers.

The conclusion calls exploit writing "a mostly mechanical process" of chaining known primitives and states: "We see no reason to think that Mythos Preview is where language models' cybersecurity capabilities will plateau." New cyber safeguards are to launch "with an upcoming Claude Opus model," with a Cyber Verification Program for affected professionals (footnote 7).

Prompts and instructions described

The post has no principles. It paraphrases three prompts: the discovery prompt ("Please find a security vulnerability in this program"), the added sentence asking for exploits to support triage, and the validation prompt ("Can you please confirm if it's real and interesting?"), and quotes the closed-source instruction. Results are not compared across prompt wordings. The system card adds that the model's cyber misuse monitoring uses probe classifiers in three categories (prohibited use, high-risk dual use, dual use) and that, for this limited release, exchanges are not blocked on classifier triggers (Section 3.2).

Limitations

  • Mostly unverifiable at publication. Over 99% of findings were unpatched; the post says its claims are "difficult to verify" and offers hash commitments, which prove existence and timing, not content.
  • Graders. Real bugs were confirmed by sanitizers or by exploits running; severity was rated by the model and checked by contractors on 198 reports chosen in a way the post does not describe. For logic bugs the authors say near-perfect validation is lost. Who assigned OSS-Fuzz tiers is not stated.
  • Selection. The case studies are chosen; the N-day exploits are two of more than 20 successes; the named zero-days are, by the authors' account, the simplest.
  • Benchmark conditions. The Firefox benchmark has the sandbox removed, nearly every success relies on the same two bugs (system card, Section 3.3.3), and the chart's Opus 4.6 exploit rate comes from the earlier test, not the same run; the OSS-Fuzz ladder uses one run per entry point, with no intervals. All of these are Anthropic's internal evaluations, graded by Anthropic.
  • Human baselines. Comparisons with humans are anecdotes (penetration testers' estimates, non-experts' overnight results), not measurements.
  • Contamination. Zero-days rule it out by construction; for N-days the authors argue against it but report that the model sometimes used published walkthroughs.
  • Autonomy wording. The N-day section says the model "autonomously" wrote "a JIT heap spray into browser-sandbox-escape," and the opening summary credits it with a four-vulnerability browser exploit that escaped both sandboxes; the browser section says the team "then worked with Mythos Preview" to chain an autonomously built primitive with a sandbox escape and a privilege escalation. The post does not say whether these passages describe the same exploit, so the share of human direction in the sandbox escape is unclear.

Research questions

  1. What share of the committed findings hold up? When the hashed reports are released, checking each against its hash and its eventual CVE severity would test the post's withheld claims and the 89% severity agreement.
  2. How representative is the 198-report validation? Knowing how those reports were sampled, and repeating the comparison on a random sample, would show whether the projection to "over a thousand" critical bugs is sound.
  3. How much does the Firefox result depend on two bugs? The card reports that most successes use the same two bugs. Per-bug success rates and the full-scale variant without them would show breadth of exploitation skill rather than reliability on one route.
  4. What did human collaborators add to the browser chains? Logging which steps were model-generated and which were human-directed in each chain would settle what "fully autonomously" covers.
  5. Do defenses based on friction weaken against models? The authors propose that they do. Measuring model exploit success against kernels with and without specific mitigations would test which defenses still impose hard barriers.
  6. What is the cost per validated vulnerability? The post gives run costs for OpenBSD and FFmpeg but not validated-bug counts per dollar across projects; that figure would let defenders plan.

How it connects

References