For years, the idea that AI would change hacking was mostly a forecast. This week the evidence moved into software makers' own security records. On 27 October Anthropic's Frontier Red Team said an unreleased internal Claude research model had built five working attack chains against widely used software. They were: a one-click attack that runs code outside Chrome's sandbox; a Safari attack that reads a user's data from other websites; taking photos off a locked Android phone; remotely crashing Windows servers that answer HTTP/3 requests; and a zero-click attack that gives full control of Linux servers running a multipath TCP listener. Anthropic says it reported every flaw to the vendor and that all of them have been fixed. [1]

Part of this can be checked independently. Google's release notes for Chrome 156 (20 October, 353 security fixes) credit several critical and high-severity bugs to "Xinyang Ge (Anthropic), assisted by Claude" and others to "OpenAI Codex Security". [2] A follow-up Chrome update a week later fixed 52 more flaws, four of them critical. SecurityWeek notes that most of the flaws that earned no bounty were found internally or with the help of AI. [3] The US National Vulnerability Database lists the Linux bug Anthropic names (CVE-2026-98432) as a kernel use-after-free that has been fixed in the stable branches. It lists the Windows bug (CVE-2026-33096) as an out-of-bounds read in HTTP.sys that lets an attacker crash a machine over the network, with a record published in April 2026. [4][5]

Other parts cannot be checked. Anthropic says the model wrote the attack chains "after initial prompting". As India's MediaNama points out, the company does not name the model, say how many attempts or how much computing it took, or report where it failed. [1][6] The public Linux record does not mention multipath TCP, so not every detail of Anthropic's account can be matched to it. Anthropic's figure of 62 Chrome CVEs is its own count. [1][4]

Attackers are using similar tools. According to Australia's iTnews, the security company CrowdStrike says a China-linked group it calls VAPOR PANDA used ARTEX, an open-source framework that connects hacking tools to models such as DeepSeek. CrowdStrike says the group targeted more than 200 organisations in under two weeks starting on 17 October and confirmed intrusions at nine. A human operator gave the instructions in Mandarin. A researcher at Google's Threat Intelligence Group called the campaign "a warning shot" and said there is "no evidence" yet that fully autonomous hacking exists. [7] The attribution comes from security companies and has not been confirmed by any government.