AI Makes More Attack Paths Affordable, but Attackers Still Seek the Least-Defended Route
Summary
The article argues that AI can lower the effort needed to exploit vulnerabilities and assess attack paths, but says evidence has not shown broad real-world acceleration. It urges defenders to consider exposure, suppliers, identities, and detection—not patch speed alone.
Key points
- In controlled tests, Anthropic’s Mythos Preview developed working exploits for some vulnerabilities in hours, including old flaws; these results do not establish broad real-world exploitation.
- A researcher’s controlled Chromium exploit experiment using Claude Opus 4.6 took about a week and required substantial human guidance; the proof of concept had limitations.
- Bitsight reported underground advertisements for AI-assisted testing, phishing, and voice-changing tools. Some seller claims, including vulnerability findings, were not independently verified.
- The article says attackers may favor overlooked routes such as stolen sessions, social engineering, exposed systems, and weaker suppliers rather than heavily monitored vulnerabilities.
- It cautions that the retired Zero Day Clock metrics could misrepresent exploitation timing and rates; available population-level data has not demonstrated broad acceleration in real-world exploitation.
- The article recommends assessing exploit likelihood alongside asset exposure and business importance, and managing exposure and detection-and-containment timelines as well as patch speed.
Article Details
- Topic
- AI-assisted vulnerability exploitation and how attackers choose less-monitored paths into organizations
MITRE ATT&CK
T1087 · Account DiscoveryAnthropic reported an increase in Account Discovery occurrences in its analysis of malicious cyber activity.T1539 · Steal Web Session CookieThe described reverse-proxy flow captures session cookies that can help an attacker access a victim's authenticated session.T1557 · Adversary-in-the-MiddleThe article describes adversary-in-the-middle reverse-proxy flows that capture credentials and session cookies during authentication.T1566.004 · Spearphishing VoiceThe article describes AI-assisted vishing and fraudulent calls intended to make victims trust the caller.
Vendors
AnthropicWhen Anthropic introduced Mythos Preview, the story practically wrote itself. Here was a frontier AI model taking work that once required researchers and threat actors a lot of time and compressing it into hours. MythosBitsightBitsight threat researchers uncovered a post on a dark web forum advertising a free automated pentesting platform with an “AI attack planner” designed to find application weaknesses and build exploitation chains. The
Products
ChromiumApril 2026, a Hacktron researcher used Claude Opus 4.6 to develop an exploit chain against an outdated Chromium release used by Discord. The experiment involved a known vulnerability in a controlled environment, andClaude CodeIn a separate campaign that Anthropic attributed to a Chinese state-sponsored group, Claude Code supported activity across multiple stages of the intrusion through an MCP-connected framework. Anthropic estimated that AIClaude Opus 4.6to see what this looks like outside a frontier-model test. In April 2026, a Hacktron researcher used Claude Opus 4.6 to develop an exploit chain against an outdated Chromium release used by Discord. The experimentDiscordresearcher used Claude Opus 4.6 to develop an exploit chain against an outdated Chromium release used by Discord. The experiment involved a known vulnerability in a controlled environment, and it required a lot ofDynamic Vulnerability Exploit (DVE) Scorerisk. Security teams cannot chase every published CVE with the same urgency. Bitsight’s Dynamic Vulnerability Exploit (DVE) Score helps narrow that enormous queue by estimating which vulnerabilities are most likely toFFmpegA 16-year-old FFmpeg flaw that caused an out-of-bounds heap write (however, Anthropic assessed that the FFmpeg flaw would be difficult to turn into a functioning exploit). Even so, the model continued to find meaningfulFreeBSDA 17-year-old FreeBSD NFS vulnerability that enabled unauthenticated root accessMicrosoft Windowsan hour and developed eight in roughly 12 hours. It also produced PoCs that triggered 18 of 21 patched Windows kernel vulnerabilities within six hours, then developed full privilege-escalation chains for eight.Mythos PreviewWhen Anthropic introduced Mythos Preview, the story practically wrote itself. Here was a frontier AI model taking work that once required researchers and threat actors a lot of time and compressing it into hours. MythosOpenBSDA 27-year-old OpenBSD bug that could remotely crash a host