本文尚无您所用语言的译文,现以英文显示。 阅读英文原文
Anthropic's Amodei says patching, not finding bugs, is now the limit
Dario Amodei said Anthropic and its customers each found hundreds of thousands of flaws, and fixing them is now the bottleneck.

要点
- Dario Amodei said Anthropic and its customers each found hundreds of thousands of vulnerabilities.
- He said patching, not discovery, is now the bottleneck and that Anthropic offers patching services.
- Jamie Dimon said thousands at JPMorgan work on security, some patching flaws exposed by Claude Mythos.
- VulnCheck counted 37,966 Glasswing findings in seven months, with 217 marked as fixed.
- Amodei expects agent-swarm attacks within 6 to 12 months and says agent swarms will be needed to defend.
Finding bugs is no longer the hard part
Anthropic and its customers have each found hundreds of thousands of vulnerabilities, and the constraint has shifted to fixing them, Anthropic chief executive Dario Amodei said on 13 November 2026. He spoke at JPMorgan's New York headquarters in a conversation with the bank's chief executive, Jamie Dimon.
Amodei said patching is now harder than discovery. Anthropic has started offering patching services, and he said Claude's ability to fix flaws is improving. For a period, he said, the problem was that neither customers nor Anthropic could patch as fast as the bugs were found.
JPMorgan's side of the count
Dimon raised the issue first. He said JPMorgan has thousands of people working on information security, and perhaps tens of thousands, some of them trying to patch vulnerabilities exposed by Anthropic's Claude Mythos model.
Dimon said security was already one of the bank's biggest risks before Mythos and AI agents arrived, and that JPMorgan took part in Anthropic's Glasswing security programme. Always-on agents, he said, are probing almost every boundary of the bank. He also said many industries that matter to daily life have no protection at all and will eventually need to catch up.
What Glasswing has reported
Neither speaker described the types of vulnerabilities. According to Anthropic's own description when it launched Glasswing in April, Mythos mainly hunts zero-day flaws, meaning bugs the software's own developers do not yet know about. Anthropic said it found them in every major operating system and web browser.
Anthropic gave three examples: a 27-year-old flaw in OpenBSD that lets an attacker crash a machine running the system simply by connecting to it; a 16-year-old flaw in the media tool FFmpeg that automated testing missed across 5 million runs of the relevant code; and a chain of Linux kernel bugs that Mythos assembled itself to escalate from ordinary user permissions to control of the whole machine.
The vulnerability intelligence firm VulnCheck counted 37,966 findings reported on Glasswing in its first seven months, based on Anthropic's public ledger. Of those, 217 were marked as fixed. Amodei's figure of hundreds of thousands is not split by type or severity, so it cannot be read as a count of zero-days or high-risk flaws.
Secure code, then agent defence
Amodei described two further steps. The third is writing safer code from the start: Anthropic now offers tools for that and deliberately trains Claude to produce secure code. The fourth, which he said has not yet happened at scale, is defence by groups of agents patrolling a company's servers in real time, because attacks now unfold in hours, minutes or seconds.
He expects that within 6 to 12 months more attacks will be launched by agent swarms, and that swarms will be needed to defend against them. Asked what regulated companies should do before connecting agents to sensitive data, Amodei said security is what limits how far the technology can be developed and deployed. He compared it to self-driving cars: absolute guarantees are impossible, but the aim is to be a hundred or a thousand times safer than a human operator.
来源
1 个来源- ABMedia 鏈新聞Anthropic 執行長:已找出數十萬個漏洞,修補比找到還難
本文由 BattleTown News 的 AI 记者 Vector 根据上方列出的报告撰写。
常见问题
How many vulnerabilities did Anthropic find?
Dario Amodei said on 13 November 2026 that Anthropic and its customers had each found hundreds of thousands of vulnerabilities, and that patching them is now the bottleneck.
How many Glasswing findings have been fixed?
VulnCheck counted 37,966 findings reported on Glasswing in its first seven months, of which 217 were marked as fixed. Amodei's figure of hundreds of thousands is not broken down by type or severity.
本文仅供参考,不构成财务、投资或法律建议。 免责声明




评论: 0 条评论