Can a language model find several vulnerabilities in heavily defended consumer software and combine them into a working attack? This post, published on 27 October 2026 by Anthropic's Frontier Red Team, answers with five worked cases. Over "the past few months," the team used "an unreleased, internal research model" to find vulnerabilities in browsers and operating systems, and the post describes five attack chains built from them (introduction). The post is the version of record: it has no companion paper, benchmark or released code, and the vendor advisories it links are the only outside records of the findings. The post is about 1,800 words long.
How the work was done, as the post describes it
Each case description contains the same phrase: "After initial prompting, Claude wrote" the exploit. The post does not publish the prompts, say how many prompting turns or how much human guidance each chain took, or say how many attempts failed. The model is not named beyond "unreleased, internal research model."
Two cases give more detail on method:
- Android. The work combined reading the open-source parts of Android, reverse engineering and deobfuscating Google Play services binaries, and live testing on a physical Google Pixel 9 (section 3).
- Windows. The team gave the model a copy of the Windows executable and asked it to reverse-engineer the software with existing tools such as Ghidra, write an annotated C program reproducing the relevant features, check the reproduction against the binary, and then look for vulnerabilities. The vulnerability surfaced when the team asked the model to identify "potentially higher-risk issues" (section 4).
Every vulnerability was reported to the vendor under Anthropic's coordinated vulnerability disclosure policy, which the post summarizes as 90 days to patch and a further 45 days before Anthropic discusses details beyond the public advisory (introduction).
The five cases
The post's descriptions are summarized here at the level of component and outcome only.
- Chrome, one click. A chain of four vulnerabilities, starting in the V8 JavaScript engine and ending outside Chrome's process sandbox, let a malicious page run arbitrary code as the user. The proof of concept ran on a stock, default-configuration Chrome on macOS "within seconds of the page loading," with no further interaction (section 1). The post adds that Google has assigned 62 CVEs from Anthropic's reports, "patched as of Chrome 156," all rated critical or high, including 13 sandbox escapes and 21 renderer-compromise vulnerabilities.
- Safari, one click. A chain through JavaScriptCore, WebKit's Gigacage isolation and the WebContent sandbox let any website's script bypass the same-origin policy and read history, cookies and responses from other sites. The post calls it a "data-only attack that never relies on arbitrary code execution," and says several of the weaknesses were "intended behaviors that become exploitable when combined." Apple acknowledged the report the same day and fixed it (section 2).
- Android lock screen. Three vulnerabilities, chained, let an attacker with a stolen phone extract photos and other personal data without unlocking it. Each of four layers of protection, from USB data blocking to user confirmation of pairing, was bypassed. Google has addressed the issue (section 3).
- Windows, remote denial of service. An out-of-bounds read in the Windows kernel's QUIC implementation let a remote attacker crash any machine serving HTTP/3 through the native Windows implementation, repeatedly. The post estimates "several million such servers" and says "The difficulty in this case was in identifying the vulnerability in the first place." Microsoft assigned CVE-2026-33096 (section 4).
- Linux, zero click. A race-condition use-after-free in the kernel networking stack let a remote attacker run commands as root on servers with a multipath TCP listener, without credentials. A lost race only fails a handshake, so the attack can be retried; the proof of concept took 48 minutes to defeat kernel address randomization and five more minutes to run a command. The bug, CVE-2026-98432, was patched upstream and backported (section 5). The same section reports, without counts, local privilege escalations in OpenBSD, FreeBSD and Linux, sandbox escapes and web-application vulnerabilities.
What the outside records show
The linked records date the work and check some counts. These checks are this note's.
- Windows. The NVD record for CVE-2026-33096 was published on 14 April 2026 and describes an out-of-bounds read in "Windows HTTP.sys" allowing denial of service over a network, scored 7.5 (High) by Microsoft.
- Safari. The linked Apple advisory is for iOS 26.5.2 and iPadOS 26.5.2, released 29 June 2026. It credits one WebKit issue, CVE-2026-43715, to "Milad Nasr and Nicholas Carlini with Claude, Anthropic."
- Chrome. The linked release notes (Chrome 156, 20 October 2026) list 353 CVEs. Fifteen are credited to Xinyang Ge, one of the post's contributors, five of them "assisted by Claude": 4 critical and 11 high. The post's 62 is therefore not a count for that release alone. This note did not trace earlier releases to confirm the total, or which entries form the described chain.
- Linux. The NVD record for CVE-2026-98432 was published on 15 October 2026 and scored 9.8 (Critical) by the assigning authority. Its description concerns an IPv4 TCP request-socket use-after-free and does not mention multipath TCP, which the post names as the listener the attack needs.
The Windows record and the Apple fix place parts of the work in April and June 2026, so, on this note's reading, "the past few months" covers at least half a year.
What the post says it learned
The post names two lessons ("What we learned"):
- Scale and scanner diversity. "Test-time compute scaling and scanner diversity substantially increase how many vulnerabilities you find." The team runs many scanning agents in parallel and measures each scanner by the additional vulnerabilities it contributes. Strategies range from per-function coverage to goal-based and threat-model-based questions, with automated variant analysis on every confirmed finding. All findings pass through a standardized verification and risk-rating pipeline. No numbers are given for any of these effects.
- Stopping new bugs. Fixing the backlog "only solves half the problem"; the post recommends scanning at pull-request time and points to a public Claude Code plugin, security-guidance, that Anthropic uses internally.
In the closing section the post states that discovering and combining vulnerabilities, once "the domain of few specialist teams," is now "something models can do with little human involvement." It points defenders to the expanded Cyber Verification Program, which gives qualifying security professionals access to Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and later models with reduced blocking classifiers.
Instructions and choices described
No prompts, principles or system text are published. The post describes three kinds of choices: the prompting phrase "after initial prompting," the Windows task decomposition quoted above, and the disclosure timeline. The Cyber Verification Program is an access policy: which users receive a model with fewer safeguards.
Limitations
- Case studies, not an evaluation. There is no task set, success rate, number of attempts, held-out set or human baseline. Cases were chosen by the authors, and the post does not say how many attempted chains failed.
- Human involvement is unmeasured. "After initial prompting" and "little human involvement" are not quantified, and the Windows case shows humans setting the decomposition of the task.
- Unnamed model. The results cannot be attributed to any released model, so they cannot be checked against a system card.
- Self-report. The post is Anthropic's account of its own model. Vendor advisories confirm that some reported vulnerabilities were real and fixed, and credit Anthropic staff, but they do not confirm the exploit chains or the model's share of the work.
- Unlocated totals. The 62 Chrome CVEs and their breakdown are not itemized in the post, and the linked release lists 15 credited to a contributor.
Research questions
- How much human steering did each chain need? Logs of the prompting turns and human interventions for each case, or a replication under a fixed prompt budget, would show what "after initial prompting" means in practice.
- What is the success rate behind the five cases? The number of targets attempted and chains completed, with time and token cost, would show whether these cases are typical or the best of many.
- How does scanner diversity trade against more compute on one scanner? The post's per-scanner "additional vulnerabilities" measure, published for a fixed budget, would test its first lesson.
- Does pull-request-time scanning lower the rate of new vulnerabilities? Comparing vulnerability introduction rates in repositories before and after adopting the security-guidance plugin would test the second lesson.
- Which released models can reproduce these chains? Running the same targets with models available through the Cyber Verification Program would connect the case studies to models outsiders can evaluate.
- How do chain-building results compare with exploit-writing benchmarks? Placing these cases beside the success rates in Anthropic's exploit-development measurements would show whether open-ended targets are harder or easier than benchmark tasks.
How it connects
- The post's opening link goes to An opt-in vulnerability-finding service for open-source software (October 2026), the open-source side of the same program.
- Partnering with Mozilla to improve Firefox's security (March 2026) reported that Claude Opus 4.6 turned a Firefox vulnerability into a working exploit in two of several hundred attempts, in a test environment with browser protections removed. This post describes full chains against default configurations; the comparison is this note's, and the two use different models and give no common measure.
- Using Claude to reduce the attack surface of code (October 2026), by one of this post's contributors, describes the defensive counterpart: removing reachable code rather than finding bugs.
- This note connects the post to Measuring LLMs’ ability to develop exploits and Assessing Claude Mythos Preview’s cybersecurity capabilities, which report exploit capability as measured results, and to Building AI for cyber defenders (September 2025), an earlier statement of the defensive rationale.
- The Cyber Verification Program is described in the Claude Fable 5.2 and Mythos 5.2 system card as one route to Mythos 5.2 for vetted users.
References
- Carlini, N., Dworken, D., Ge, X., Morris, B., Nasr, M., Rohlf, C., & Tong, L. (2026, October 27). Discovering and exploiting vulnerabilities in critical, real-world software with Claude. Anthropic Frontier Red Team. https://www.anthropic.com/research/discovering-vulnerabilities-with-claude record
- Google. (2026, October 20). Stable Channel Update for Desktop (Chrome 156). Chrome Releases. https://chromereleases.googleblog.com/2026/10/stable-channel-update-for-desktop_02018666648.html record
- Apple. (2026, June 29). About the security content of iOS 26.5.2 and iPadOS 26.5.2. https://support.apple.com/en-us/127594 record
- National Vulnerability Database. CVE-2026-33096 (published 2026-04-14) and CVE-2026-98432 (published 2026-10-15). https://nvd.nist.gov/vuln/detail/cve-2026-33096 and https://nvd.nist.gov/vuln/detail/cve-2026-98432 record
- Linked from the post and not read for this note: Anthropic's Coordinated Vulnerability Disclosure policy, the Cyber Verification Program announcement, and the security-guidance plugin repository.