Palo Alto Networks finds ungoverned AI at 99% of enterprises; Prisma AIRS passes $100 million in ARR
The sixth State of Cloud Security Report, from 2,800 security executives, is the clearest demand document yet for agent security, published two days before Palo Alto reports earnings with its AI security product already past $100 million of ARR. CrowdStrike is selling a kill switch into the same budget.
Richard Tangthree-minute read
Enterprises have handed AI agents the keys to production faster than they built any way of seeing what those agents do. Palo Alto Networks−2.90% — Palo Alto Networks, down 2.90 percent today, the cybersecurity vendor, published its sixth State of Cloud Security Report on November 16, a survey of 2,800 security executives and practitioners across 10 countries, and it is the clearest demand document yet for agent security: 99% of organizations found ungoverned AI in their environment in the past year, and 9% can discover it automatically 12. The binding constraint on enterprise AI is no longer model capability. It is plumbing: identity, audit and revocation.
Agents act on borrowed names, and the audit trail breaks
Agents can modify code, pipeline configurations or production environments at 73% of organizations, and at 53% of them engineers log agent actions under their own identity or personal access token 12. Enterprise security assumes a person stands behind every credential. Break that assumption and 40% of organizations cannot reconstruct what an agent did after the fact 12. Frontier models can chain flaws into working exploits in minutes, while the average fix takes 18.4 days to reach production 31.
Confidence is near-universal; the ability to see is not
Almost every respondent expresses confidence in securing the AI supply chain, while the capabilities that confidence would rest on, watching agents in production, catching prompt injection, finding ungoverned AI at all, sit with a small minority 12.
Confidence runs far ahead of the ability to see AI at work
97% express confidence; only 9% can automatically find ungoverned AI
Data
| Value | |
|---|---|
| Confident in AI supply-chain security | 97% |
| Can monitor AI behavior in production | 36% |
| Can detect prompt injection | 30% |
| Can auto-discover ungoverned AI | 9% |
The diagnosis doubles as the sales pitch
Palo Alto sells the cure it describes. Prisma AIRS, its AI security product, passed $100 million of annual recurring revenue on 800 customers, in a line guided to 22–23% ARR growth this fiscal year, and the company reports earnings on November 18 4.
Revenue growth reaccelerated to 34.5% in Palo Alto Networks' latest quarter
Data
| Revenue | |
|---|---|
| Q3 '23 | $1.88B |
| Q4 '23 | $1.98B |
| Q1 '24 | $1.99B |
| Q2 '24 | $2.19B |
| Q3 '24 | $2.14B |
| Q4 '24 | $2.26B |
| Q1 '25 | $2.29B |
| Q2 '25 | $2.54B |
| Q3 '25 | $2.47B |
| Q4 '25 | $2.59B |
| Q1 '26 | $3B |
| Q2 '26 | $3.41B |
Eight hundred customers against a problem present almost everywhere says the product is early, not late. CrowdStrike−1.96% — CrowdStrike, down 1.96 percent today wants the same budget: its AI Kill Switch, launched November 3 inside the Falcon Guardian platform, lets a human terminate a rogue agent at the root 5.
On November 17, Hitachi Systems India opened a Bengaluru delivery center to sell that stack, with QuiltWorks defense services built on OpenAI$1.18T — OpenAI, private, latest valuation $1.18T, Anthropic$350B — Anthropic, private, latest valuation $350B and Nvidia−1.97% — Nvidia, down 1.97 percent today models, across India, Asia-Pacific and the Middle East 6. CrowdStrike reports on December 1 4.
Both defenders hold a seat at the gated frontier
The survey's own questionnaire grades the threat "Mythos-grade," after Anthropic's Claude Mythos models, which Anthropic keeps out of general release 1. Palo Alto and CrowdStrike were both launch partners in Project Glasswing, the Anthropic program that in April gave 12 defenders early access to Mythos Preview, which Anthropic says has surfaced thousands of critical zero-day vulnerabilities 7.
On the other side of the Pacific, Alibaba−3.93% — Alibaba, down 3.93 percent today, Tencent and Moonshot are stripping offensive cyber capability from their newest models ahead of a draft Chinese export ban 8. The strongest attack models are being gated everywhere, and the vendors holding defensive seats are the ones selling the controls. That access, more than any feature list, is the moat.
Visibility sells; the rest has to be rebuilt
Discovery and monitoring will convert to billings quickly, because an enterprise can purchase the ability to see. The rest of the gap is process: a fix cycle measured in weeks, revocation that still needs a person, and nearly half of teams correlating vulnerabilities with live threats by hand 1.
That work happens inside each organization. Wendi Whitmore, Palo Alto's chief security intelligence officer, states the conclusion correctly: trust in AI "will be determined by the cybersecurity controls we build around" the models 2. The agents should keep running. The value being created now sits with whoever gives each agent its own identity, a clear owner and a switch that works.
Get Security by email, free
The breaches, flaws and security vendors AI is putting to the test. In our daily email, every weekday.






Reader comments
0 comments