Microsoft's 2026 Digital Defense Report: Median Time From Vulnerability Discovery to Weaponization Is Now Well Under 24 Hours, and Frontier Models Chained 32 Attack Steps on Their Own

On October 1 Microsoft released its 2026 Digital Defense Report, covering July 2025 to June 2026. Its central finding is that AI has given attackers a temporary lead: the median time from discovering a vulnerability to weaponizing it is now "well below 24 hours," and CVEs registered in 2026 are on track for a record of about 72,000. Among the intrusions Microsoft investigated, phishing as the entry point rose from 7% to 23%, and exploitation of public-facing applications from 15% to 24%. In a controlled evaluation, Anthropic's Mythos Preview and OpenAI's GPT-5.5 chained 32 attack steps without human direction to take over an entire domain in an emulated enterprise network. The report also notes that target selection and key decisions are still made by humans in most attacks.

The number that stands out: less than a day

Microsoft's annual report covers July 2025 to June 2026 and opens with the line "AI is changing the physics of cybersecurity." The clearest evidence is time: the median gap between a vulnerability being discovered and being turned into a working attack is now "well below 24 hours." That step used to take senior researchers; in many cases it now comes down to "simply writing a prompt." Meanwhile, CVEs registered in 2026 are on track for a record of about 72,000.

How attackers get in is shifting too. Among the intrusions Microsoft investigated, phishing as the entry point rose from 7% to 23%, and exploitation of public-facing applications from 15% to 24%. The report's explanation is that the old tells for spotting phishing, forged documents, sloppy writing, an accent on video and a thin online footprint, can all be fixed by AI at once.

Models running a full attack chain by themselves

In a controlled evaluation, Anthropic's Mythos Preview and OpenAI's GPT-5.5 executed 32 consecutive attack steps in an emulated enterprise network with no human direction, ending with full domain compromise. The report also describes the s1ngularity malware from August 2025: spread through trojanized Nx npm packages, it looked for AI command-line tools such as Claude Code and Gemini CLI on infected machines and ran them with permissive flags to hunt for secrets, leaking about 2,000 secrets and 20,000 files from 225 victims.

What it means

The report adds a sober footnote: in most attacks, target selection, operational decisions and the most complex intrusion steps are still done by people. For defenders, its advice is concrete: make phishing-resistant multi-factor authentication and passkeys the baseline; patch internet-facing systems on the assumption that flaws will be exploited within a day; and hold AI agents to the same least-privilege standard as human accounts. That last point is the most practical one for developers: the coding agent installed on your own machine is itself a tool attackers want to borrow.

via: Microsoft 2026 Digital Defense Report, Help Net Security report, BleepingComputer report