---
title: "Weekly Roundup: Two labs published exploit counts, and an agent talked analysts into its own coin flip"
pubDatetime: 2026-11-06T15:00:00.000Z
description: "Anthropic's research model produced 62 Chrome CVEs, OpenAI's pipeline found 948 kernel root exploits for $1.2M in credits, and ARTEX showed up in four separate posts."
tags: [ai-security, exploit-development, ai-agents, anthropic, openai, vulnerability, 2026, 2026-q4, 2026-11, weekly-roundup]
---
> [!tldr] TL;DR
> Both frontier labs spent the week publishing how many working exploits their models found, and the numbers are large.

Anthropic described [five working attack chains](/posts/anthropic-five-attack-chains-claude) built by an unreleased research model against Chrome, Safari, Android, Windows and Linux. The Chrome work alone produced 62 critical or high CVEs, all patched by Chrome 156. A week later OpenAI reported [948 Linux kernel privilege-escalation exploits](/posts/openai-948-kernel-lpe) from a single mid-October run, at a cost of $1.2M in API credits. Ten days after the run, 61 were patched upstream. That ratio is the part worth sitting with.

The other Anthropic result got less attention and deserves more. Fifty expert analysts were given a Claude Code agent whose position on every judgment call was [set by a secret coin flip](/posts/anthropic-agents-judgment-calls). They went along with it 86% of the time, and most reported the conclusion as their own.

ARTEX turned up in four posts: [espionage across 200+ targets in 11 days](/posts/auto-artex-ai-espionage), [agents making their own exploitation decisions](/posts/auto-artex-autonomous), the [WorkOS breach](/posts/auto-artex-commodity), and a [backdoored Docker image on 7,946 instances](/posts/auto-artex-backdoor). An open-source pentest framework becomes attacker infrastructure, then gets supply-chained itself.

Also this week: ChatGPT's Lockdown Mode had [an iframe sandbox escape](/posts/auto-chatgpt-lockdown-bypass), Anthropic's [FlashProve](/posts/anthropic-flashprove-provable-inference) proves which weights answered at 70-95x the GPU cost, and GSA [cut 40% off Claude](/posts/anthropic-gsa-government-discount) one day after DOJ moved to restore a federal ban on it.