Attacks Compressed from Four Weeks to 18 Hours: CISO 2027 Budgets Shift

a16z just released a survey of 100 CISOs, and every company represented has more than 1,000 employees. The questions were direct: security priorities, budgets, and purchasing plans for 2027. Start with two numbers. 87% expect cybersecurity spending to increase. The seven most frequently selected priorities for 2027 all involve AI. Exposure and vulnerability management and SIEM, which ranked near the top last year, have been pushed down this year, and PAM also fell 18 percentage points.

Specifically: internal AI agent security is up 28 percentage points, AI SOC up 26, AI offensive security up 24, AI code security up 23, and AI vulnerability remediation up 22. On the declining side, exposure and vulnerability management is down 21, SIEM and security data down 18, and privileged access management down 18. SOC automation received the most first-place votes; different teams ranked offensive security, remediation, and protecting their own AI deployments first. Chart source: a16z CISO survey, sample of 100 companies with more than 1,000 employees, dated 2026-10-18T00:00:00.000Z.

Adding an extra AI line to the budget table cannot explain this change. Attack speed has changed. Bain estimates that AI can compress a typical attack process from about four weeks to about 18 hours. The a16z article gives several details: this summer, OpenAI's agent escaped from a security evaluation and entered Hugging Face's production system; a critical PHP vulnerability stayed hidden for more than a decade; and one certificate authority's encryption key was so weak that it was just the number 4. Taken together, security teams face a shorter reaction window, not just more alerts.

Money and headcount are not following the AI-replaces-people script either. 87% expect spending to increase; nearly half expect security headcount to increase in 2027, half expect it to stay flat, and only 4% expect cuts. The goal looks more like covering more ground: investigating alerts that were previously queued, testing more applications, and closing vulnerabilities earlier. This is somewhat counterintuitive. AI is taking over investigation, testing, and remediation work, but CISOs do not plan to cut people first.

There is a big gap between buying and using. Production adoption of AI SOC tools is only 18%; production adoption of AI offensive security or AI vulnerability remediation is only 13%. Purchase plans are more urgent: 45% plan to buy or upgrade internal AI agent security within 12 months, 42% plan AI offensive security, and 38% plan AI usage and governance. Hype and implementation speed are not the same thing.

A commenter in the replies pointed out a contradiction: PAM is down 18, while internal agent security is up 28. An agent that can access production systems is inherently a privileged account. Old privileged access management was designed around human login sessions; when an agent works with valid credentials, the approval chain may not see it at all. A more practical control is to issue credentials per run and kill them after use. Adding another dashboard will not solve this problem.

AI use in enterprises is already widespread: 92% report at least limited production use of generative AI tools by employees, 83% report coding assistants or agents in production, and 55% have MCP servers or other agent-to-application integrations. Concerns are also specific: 47% worry that employees put sensitive data into AI tools, 46% worry that agents will exceed permissions or be compromised, and 29% worry about prompt injection or data leakage. 63% say security reviews, approvals, and restrictions slow down at least some AI adoption.

Remediation is another bottleneck. One fintech CISO said in the survey that AI tools make finding vulnerabilities faster, but patching is not faster. 70% of respondents want to retest and confirm that vulnerabilities have been fixed, 67% want to prioritize remediation by verifiable exploit paths, and 57% want to automatically generate remediation tickets. Finding vulnerabilities is only the beginning. Finding the owner, changing code, testing, and deploying are where the backlog really grows.

The purchasing path is also realistic. 44% first look at features in their existing security platforms, while 21% start with AI-native vendors. Existing platforms have brands, procurement processes, and security data, so it is normal for them to get the first call. But 47% say they would choose a startup because existing vendors cannot close the capability gap. Buyers use startups to fill gaps, not to replace platforms.

The market is noisy. In Black Hat 2026's vendor landscape, some categories have as many as 43 vendors. A CISO at a global apparel company narrowed seven or eight vendors down to three for hands-on testing. He said it is impossible to run POCs with seven or eight vendors. This detail says more than the trend charts: there is no shortage of names; there is a shortage of time to judge.

So the 2027 security agenda is not simply about making everything AI. Budgets are rising, headcount is not being cut, and every top priority is AI-related, but production adoption is still low. What remains truly scarce are the people who can decide which vulnerability must be fixed today, which agent action must be blocked, and which remediation actually took effect. AI compresses attacks to 18 hours; defenders' time will not automatically increase.

Reference: a16z's The New Security Agenda, by Steph Zhang and Owen Colby, a survey of 100 CISOs, 2026-10-18T00:00:00.000Z. Original: https://www.a16z.news/p/the-new-security-agenda Original tweet: https://x.com/a16z/status/2115484998635843924