Anthropic’s Frontier Red Team has revealed that an unreleased internal Claude research model generated complete, multi-step exploit chains for five major software platforms — Chrome, Safari, Android, Windows, and Linux, after a single prompt. The model didn’t just flag individual bugs. It identified separate vulnerabilities across different components, then linked them into functional attack paths.
Each chain bypassed multiple security layers: four in Chrome, four in Safari, four in Android, four in Windows, and four in Linux. Anthropic says the model performed reverse engineering, deobfuscation of binaries, and live testing on real hardware, including a Pixel 9, to build working proofs of concept. All exploits were reported to vendors and patched before public disclosure.
None appeared in any released version of Claude. No user-facing product, free or Pro, is affected. The company stresses this was red-team research, not a demonstration of deployed capability. It does not indicate that current Claude models can replicate these results.
The team did not name the internal model, nor did it say whether the exploit-generation ability exists in Claude 4, Sonnet 4.5, or the upcoming Opus 5.5. Anthropic confirmed it shared findings with Google, Apple, Microsoft, and Linux kernel maintainers under coordinated disclosure. Google assigned 62 CVEs to the Chrome issues alone, all rated critical or high, and patched them in Chrome 156.
Thirteen involved sandbox escapes. 21 compromised the renderer process. Apple acknowledged Anthropic’s Safari report the same day it was submitted and shipped a fix that disables the same-origin policy bypass, a data-only attack that could read browsing history, cookies, and authenticated session responses without code execution. On Android, the model extracted photos and personal data from a locked Pixel 9 by combining flaws in open-source Android code and proprietary Google Play services binaries.
Microsoft addressed its vulnerability, an out-of-bounds read in the Windows kernel’s QUIC implementation, as CVE-2026-33096. Anthropic estimates several million Windows servers exposed to the internet remain vulnerable unless updated. The Linux exploit achieved zero-click root access via multipath TCP (CVE-2026-98432), a race-condition-induced use-after-free flaw. The proof of concept took 48 minutes to leak a kernel pointer, then five more to execute arbitrary commands.
The team also found local privilege escalation flaws in OpenBSD and FreeBSD, plus sandbox escapes in widely used infrastructure, but only disclosed technical details for the multipath TCP and Chrome sandbox cases. Anthropic’s post notes gaps in publicly available information, not gaps in its own findings. In India, no CERT-In advisory has been issued.
There is no indication that Indian users face new risk from using Claude AI via web or app. The service remains free in India. The Pro tier costs ₹1,299 per month or ₹11,999 annually, identical to US pricing ($19.99/month or $199/year). No regional rollout, feature update or pricing change is tied to this disclosure.
TBFiled by The Tech Business Desk
Big tech, startups, earnings, regulation and policy — the decisions that shape the products.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.