N019: Anthropic's unreleased model chained four Chrome flaws into a one-click exploit
Anthropic's unreleased model built a one-click Chrome exploit and a no-login Linux root exploit, while attackers ran a 44-agent kit and an AI-driven intrusion hit 123 Reg.
N019 · Thursday 29 October 2026 · 5-min read · Covers 28 Oct 11:30 to 29 Oct 11:30 UTC
Jump to: Incidents · Agent threats & vulnerabilities · Models & infrastructure · Autonomy, swarms & robotics · Defenses, tools & GitHub · Other big security news · Pro deep dives
Today in 30 seconds
- Anthropic says an unreleased model chained four Chrome flaws into a one-click sandbox escape and built a no-login Linux root exploit; all are patched.
- TrendAI found attackers using CyberStrikeAI, an open-source kit that splits a break-in across 44 AI agents, against at least 19 organizations, 16 in Japan.
- UK registrar 123 Reg says an AI-driven intrusion read customer data through an internal API, using copies of agent skills Anthropic had published.
Top story
Anthropic's unreleased model chained four Chrome flaws into a one-click exploit
27 Oct 2026 · Anthropic Frontier Red Team · Research · 62 Chrome CVEs fixed by Chrome 156 · CVE-2026-98432 · CVE-2026-33096 · Relevance 9 · Capability shift
An AI model found bugs in a browser and in Linux and wrote working attacks from them, work that took expert teams weeks. After "initial prompting", Anthropic's unreleased model chained V8, sandbox-escape, CacheStorage and FedCM flaws so one page visit ran code outside Chrome's sandbox, and built a network root exploit for Linux multipath TCP (CVE-2026-98432). All five chains are patched; capability claims are Anthropic's own.
New risk for leaders: patch-to-exploit time now depends on model time, not expert time; patching still works, but only if it lands in days.
For leaders: Set a patch deadline in days for browsers and internet-facing Linux, and ask key vendors whether they use frontier models to find their own flaws first.
For builders: Get every managed Chrome to 156+, patch kernels for CVE-2026-98432, and set net.mptcp.enabled=0 where multipath TCP is unused.
Incidents
123 Reg says an AI-driven intrusion read customer account data
28 Oct 2026 · The Register, quoting 123 Reg's statement and customer email · Incident · UK · Relevance 8
A UK domain-name company says attackers used AI tools to get in and read customers' contact and account details. GoDaddy-owned 123 Reg says the 18 September intrusion read names, phone numbers and product data through an internal API, and attacker servers held copies of skills Anthropic open-sourced; the China link is its own assessment.
What it means for US companies: 123 Reg notified the FTC, and registrar data is ideal bait for phishing whoever controls your domains.
Do this: Turn on two-step verification and registrar lock for every company domain.
Agent threats & vulnerabilities
An AI agent found four flaws in particle-accelerator control software in under four hours
27 Oct 2026 · Causal Security · CISA ICSA-26-300-01 · CVE-2026-114595 (CVSS 9.8), CVE-2026-114596, CVE-2026-114597, CVE-2026-114598 · Relevance 7
An AI agent read unfamiliar industrial control code and found bugs that could crash or take over lab controllers. Causal Security's Cipher agent found four flaws in EPICS Base's Channel Access and pvAccess protocols in 3 hours 49 minutes, with one human instruction; 7.0.11 fixes them.
Do this: Upgrade to EPICS Base 7.0.11 and block ports 5064-5065 and 5075-5076 at the control-network edge.
Models & infrastructure
Also:
- Moonshot's Kimi K3.1 Pro left cyber data out of training: hacking score fell 90.5% to 81.0% (self-reported); weights due 4 Nov.
Autonomy, swarms & robotics
Attackers split break-ins across 44 AI agents and hit at least 16 Japanese organizations
27 Oct 2026 · The Hacker News, reporting TrendAI research (secondary) · Campaign · Japan, Taiwan, Hong Kong · Relevance 8 · Capability shift
Attackers used a free tool that hands each step of a hack to a different AI agent, so one crew could probe many companies at once, in fluent Japanese. TrendAI says CyberStrikeAI's 44 agents ran reconnaissance through data theft and built Japanese fake login pages; 19 organizations were probed.
New risk for leaders: one operator can run many parallel, localized intrusions; patching and MFA still work, but old weaknesses get found faster.
What it means for US companies: the kit is public, and language no longer protects any market.
Do this: Close default credentials and exposed admin pages; hunt for one source probing many hosts.
Defenses, tools & GitHub
Cisco open-sources Project Quarry, rules that flag risky sequences of agent actions
28 Oct 2026 · Cisco · cisco-ai-defense/project-quarry · Apache-2.0 · new repo (created 27 Oct, 6 stars) · Relevance 7
Cisco released a free tool that flags combinations of agent actions that look like theft, even when each step looks harmless. YARA-like rules link events such as a sensitive file read followed by an outside send, with plugins for Claude Code, Codex and MCP agents.
Do this: Run it on a week of stored agent traces and tune before enforcing.
Other big security news
- Oracle PeopleSoft: ShinyHunters exploited three chained flaws, now patched as CVE-2026-35273; owners should patch and hunt for web shells (actively exploited).
- Adobe Commerce: CISA added PolyShell, CVE-2026-48356 (CVSS 9.3), to KEV after hundreds of stores were breached; Magento operators should patch now (actively exploited).
- Sony So-net and Plala: About 2.58 million customer records were exposed, per Help Net Security (secondary).
One concept
Exploit chain
Bugs used in sequence, each unlocking the next: run code in a web page, escape the browser sandbox, take over the machine. Several "high" bugs can add up to a full takeover.
Basic mitigation: patch browser and kernel bugs fast even below "critical", and keep layered sandboxes.
AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.
Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup