Anthropic's Five Exploit Chains: The Patch Checklist First, Then the AI Story

Fixed versions for Chrome 156, Linux CVE-2026-98432 and Windows CVE-2026-33096 from Anthropic's exploit-chain post, and where the CVE records disagree.

Anthropic’s Frontier Red Team published an account on October 27 of five working exploit chains built by what it calls an “unreleased, internal research model”. The five are a one-click Chrome chain in four stages, a one-click Safari chain, a Pixel 9 lock-screen bypass, a Windows remote denial of service and a zero-click Linux kernel bug. Two carry CVE IDs in the post, CVE-2026-33096 for Windows and CVE-2026-98432 for Linux. The post never names the model, and we are not going to guess. S5 Labs has not seen the exploits, the model or any proof of concept; the fixed versions below come from vendor records, which in two cases describe the bug differently from Anthropic. Anthropic, Frontier Red Team post

The patch checklist

TargetCVEComponent (vendor record / Anthropic)Fixed inSource
Chrome desktop62 CVEs, Anthropic’s count; no listRenderer and sandbox, per Anthropic156 stable, 156.0.8078.47 (Oct 20); a second update, 156.0.8078.73/.74 (Oct 27), adds more Anthropic-credited fixesChrome Releases, Oct 20 · Oct 27
Linux kernelCVE-2026-98432IPv4 TCP request sockets / “multipath TCP listener”7.3; backports 7.2.10, 6.18.56, 6.12.113, 6.6.159, 6.1.190, 5.15.223, 5.10.272CVE.org · kernel.org
Windows Server and clientCVE-2026-33096Windows HTTP.sys / Windows kernel QUIC, HTTP/3April 14, 2026 cumulative updates (builds below) or laterMSRC
Safari, iOS, macOSNone given by Anthropic; post links Apple’s CVE-2026-43715 pageWebKit (Apple) / Safari same-origin policy (Anthropic)iOS and iPadOS 26.5.2 for CVE-2026-43715; current Apple releases are from Oct 26Apple 127594 · Apple releases
Pixel 9 (Android)None givenLock screen, with a USB fallback, per Anthropic“Addressed by Google”; no version or bulletin entryAnthropic post

Neither CVE is in CISA’s Known Exploited Vulnerabilities catalog as of the October 28 feed, and Microsoft marks CVE-2026-33096 as not publicly disclosed and not exploited. CISA KEV · MSRC

Windows: Microsoft says HTTP.sys, Anthropic says QUIC

Microsoft and the CVE text describe CVE-2026-33096 as an out-of-bounds read (CWE-125) in Windows HTTP.sys: Important, denial of service, CVSS 3.1 7.5, published April 14, 2026, exploitation “Less Likely”. QUIC and HTTP/3 do not appear in the CVE description. Anthropic describes the same ID as a bug in the Windows kernel’s QUIC implementation, reachable over HTTP/3, and attaches a “several million” estimate to it; the post, not Microsoft, is the source of that number. Both can be true, since HTTP.sys is where Windows serves HTTP/3 and Microsoft’s mitigation guidance does say HTTP/3, but the CVE record alone would not tell you this is a QUIC bug. NVD · MSRC · Anthropic post

The firmest link between the two is Microsoft’s April CVRF document, which credits “Milad Nasr (Anthropic) and Calif.io with Claude” alongside Microsoft’s WARP and MORSE teams. Nasr is a named contributor on Anthropic’s post. The NVD and CVE.org records carry no credit. MSRC CVRF, April 2026

Windows versionFixed buildKB
Windows 11 23H210.0.22631.6936KB5082052
Windows 11 24H210.0.26100.8246KB5083769
Windows 11 25H210.0.26200.8246KB5083769
Windows 11 26H110.0.28000.1836KB5083768
Windows Server 202210.0.20348.5020KB5082142
Windows Server 2022 23H2 (Server Core)10.0.25398.2274KB5082060
Windows Server 202510.0.26100.32690KB5082063

Builds from the MSRC CVRF; later cumulative updates supersede them. Windows 10, Server 2019 and Server 2016 are not on the affected list. The mitigation, for machines that cannot take the update, is to back up and remove the EnableHttp3 and EnableAltSvc values under HKLM\SYSTEM\CurrentControlSet\Services\HTTP\Parameters and restart; every http.sys-based server on that machine then stops serving HTTP/3. Microsoft does not say whether HTTP/3 is on by default for a given SKU. MSRC CVRF · MSRC

Linux: the record says IPv4 TCP, the post says multipath TCP

CVE-2026-98432 was published by the Linux CNA on October 15 at CVSS 3.1 9.8. The record and the fixing commit describe ordinary IPv4 TCP request-socket handling, tcp_v4_syn_recv_sock and IP options in the SYN-ACK, fixed with kfree_rcu() in inet_sock_destruct() and tcp_v4_reqsk_destructor(). IPv6 is not affected by the first issue; a second path needs net.ipv4.tcp_migrate_req=1 or a BPF_SK_REUSEPORT_SELECT_OR_MIGRATE program. Neither the record nor the commit mentions MPTCP. Anthropic calls the target a “multipath TCP listener”, and no source we found connects the two descriptions. If the public record is right, any IPv4 TCP listener is exposed, so do not gate the patch on net.mptcp.enabled. CVE.org · NVD · Anthropic post

The fix is commit e3f33b0a by Eric Dumazet, committed October 5 with a Reported-by line for Xinyang Ge at anthropic.com. It landed in 7.3-rc7 and shipped in Linux 7.3 on October 18; we read it on the GitHub mirror because git.kernel.org was serving a bot check. CVE.org lists the kernel as affected from 4.4 and fixed at 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10 and 7.3. There is no 5.4 or 4.19 entry, and end-of-life trees get no fix. Commit on GitHub · CVE.org · kernel.org

Distribution status on October 28: Debian has fixes in bookworm-security (6.1.190-1, DLA-4841-1), trixie-security (6.12.113-1, DSA-6568-1) and the bookworm linux-6.12 backport (6.12.113-1~deb12u1, DLA-4850-1). Ubuntu rates it high but lists jammy, noble, 26.04 and 26.10 as “pending”, with no released kernel yet, and bionic, focal and xenial as “needed”. Red Hat rates it Moderate at CVSS 5.9, high attack complexity, availability impact only, and labels those metrics preliminary; we did not verify its RHEL version list. The gap between 9.8 and 5.9 has not been settled. Debian tracker · Ubuntu · Red Hat

Chrome: two stable updates, and what your headless bundles pin

Chrome 156 reached stable on October 20 as 156.0.8078.47 (Linux) and .47/.48 (Windows, Mac) with 353 security fixes; 15 credit Xinyang Ge or Anthropic, four Critical and 11 High, five of them “assisted by Claude”. A second stable update on October 27, 156.0.8078.73 (Linux) and .73/.74 (Windows, Mac), lists 52 fixes, four Critical, 46 High and two Medium. Eighteen credit Ge or Anthropic, including all four Criticals: CVE-2026-118306 (FileSystem), CVE-2026-118331 (FileAPI), CVE-2026-118348 (Skia) and CVE-2026-118344 (Layout). It is rolling out “over the coming days/weeks”. The 62-CVE figure is Anthropic’s; Google publishes no per-reporter total, and the notes do not say which CVEs belong to the chain. Chrome Releases, Oct 20 · Chrome Releases, Oct 27 · Anthropic post

Browser-agent fleets are the exposure that matters for practitioners: a one-click chain needs only a page load, and a headless agent loads whatever page it is told to. Both major automation libraries pin a Chromium older than the October 27 line. Puppeteer v25.16.0 (October 22) pins Chrome and chrome-headless-shell at 156.0.8078.47; its main branch rolled to .73 on October 28, unreleased. Playwright v1.64.1 (October 14) pins chromium and chromium-headless-shell at 156.0.8078.4, a pre-stable build; its main branch is on Chromium 157.0.8103.15, also unreleased, and 157 is not the October 27 fix line. We did not check whether any chain bug separates .47 from .73, so treat .73 as the target. Puppeteer v25.16.0 · Playwright v1.64.1

Apple and Android: nothing to check against

Anthropic gives no CVE or fixed version for the Safari chain, which it describes as a data-only same-origin-policy bypass. The post links Apple’s page for iOS and iPadOS 26.5.2 (June 29, 2026), where a WebKit use-after-free, CVE-2026-43715, is credited to “Milad Nasr and Nicholas Carlini with Claude, Anthropic”. Apple says WebKit, not JavaScriptCore, and nothing about the same-origin policy; the link between that CVE and the chain is Anthropic’s, and we cannot confirm it. Apple shipped iOS and iPadOS 27.2, iOS 26.7.2, macOS 27.2, 26.7.2 and 15.8.2, and Safari 27.2 on October 26; we have not read those pages’ credits. Read iOS 26.5.2 as the floor and the current release as the target. Apple 127594 · Apple security releases · Anthropic post

For the Pixel 9 chain, three bugs plus a raw-byte USB fallback that Anthropic tested on a physical device, the only patch status is Anthropic’s line that “this issue has been addressed by Google.” The Android bulletins for August through October and the Pixel bulletins for September and October carry no Anthropic credit, so there is no version to write down. Anthropic post

The model, the program and the plugin

The post says the five chains are a selection. The model also found unnamed local privilege escalations on OpenBSD, FreeBSD and Linux, sandbox escapes, web-application bugs and, in the post’s plural, zero-click Linux kernel networking vulnerabilities, of which one is the example written up. The contributors are Carlini, Dworken, Ge, Morris, Nasr, Rohlf and Tong. This is a different channel from OSS Scanner, which emails model-written reports to enrolled maintainers without review. Anthropic post

Anthropic also restates the Cyber Verification Program expansion from October 6: three tiers covering Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and new models going forward. The October 6 page lists Opus 5.5, Fable 5.1 and Sonnet 5.5 as generally available with conservative cyber safeguards. It limits Specialized Access to a small set of verified organizations, keeps Red Team Access to organizations, and still blocks actions that could cause physical harm or mass disruption in real time. Our CVP tiers article covers who should apply for which; the Mythos and Project Glasswing piece has the background, and Google’s Gemini 4 Carbon is taking its own defenders-first route. Anthropic, CVP expansion

The Claude Code security-guidance plugin is now public in the official plugin repository. Its README describes three layers: regex reminders on Edit and Write for about 25 patterns, an LLM diff review, and an agentic commit review that runs on git commit; the post separately recommends scanning at PR time. It installs with /plugin install security-guidance@claude-plugins-official, needs Claude Code 2.1.144 or later and Python 3.8+, and was bumped to 2.0.17 on October 27. The review model is the newest Opus the configured endpoint serves, falling back to claude-opus-5-5; SECURITY_REVIEW_MODEL pins it. We did not re-verify the README’s cost figure, its environment switches or its statement about what is sent to the model. security-guidance plugin

The disclosure terms

Anthropic’s coordinated disclosure policy, last updated March 6, 2026, sets a 90-day deadline and publishes after 90 days or a patch, whichever comes first, with a 14-day extension available. Actively exploited critical bugs get a seven-day target for a patch or mitigation, extendable by seven more days only if the maintainer is actively working on a fix. Thirty days of silence escalates to an external coordinator, and model-originated reports are labeled as such. The policy also says Anthropic “would generally wait 45 days” after a patch before publishing full technical details, while the post frames them as following the 90; plan around the policy’s version. Anthropic, coordinated vulnerability disclosure · Anthropic post

What to check

Headless Chrome. Run chrome --version on whatever binary your agents launch, chrome-headless-shell included, and compare it with 156.0.8078.73. Puppeteer’s and Playwright’s bundled browsers are on .47 and .4 until their next releases; pointing executablePath at a system Chrome on .73 gets there sooner.

Linux. Compare uname -r with the fixed list, then check your distribution’s tracker rather than kernel.org, since Ubuntu’s fixes were still pending on October 28. A disabled MPTCP is not protection; the public record describes plain IPv4 TCP.

Windows. Confirm the build against the table. On servers that must stay unpatched, decide whether losing HTTP/3 is acceptable before removing the two registry values; the change applies to every http.sys application on the host.

Apple. Install the October 26 releases. A device pinned below iOS 26.5.2 predates the one Anthropic-credited WebKit fix we could find.

Android. Nothing version-specific to verify; keep the Pixel security patch level current and watch the Pixel bulletins.

Sources

Anthropic

Windows CVE-2026-33096

Linux CVE-2026-98432

Chrome and browser automation

Apple

Continue reading.

Insight9 min read

Anthropic Flipped Claude Code to Auto Mode Because You Were Approving 97% of Prompts Anyway

Claude Code auto mode is now the default on Pro, Max and Team. Its study says humans caught 13.6% of bad commands. What to audit before trusting it.

Insight10 min read

Three Labs' Models Reached Outside Their Test Harness in 15 Days. The Sandbox Was the Bug Each Time.

AISI logged 19 unsanctioned agent actions in a cyber eval; Meta's Muse Spark 1.1 hit a third-party service. Four disclosures, one failure: sandbox egress.

Insight10 min read

Gemini Computer Use Is GA: A Builder's Checklist Against GPT-6.1 Terra

Gemini computer_use is GA. Which model to pick, three environments, the intent field, two safety traps, and what it costs next to GPT-6.1 Terra.