Trust in AI is sliding, and AI is now the first suspect whenever something big breaks.
When the FAA halted departures nationwide on Nov. 13, speculation landed fast on SMART, an AI-assisted airspace planning system the agency had expanded to all 22 air traffic control centers three days earlier. The FAA says the cause was a technical problem in its aging communications network, not AI. A day earlier, a Microsoft 365 outage knocked out Outlook, Teams and Copilot worldwide. Add the local fights over AI data centers and widespread doubt that anyone is regulating the technology, and AI sits in the hot seat whether or not it earned the spot.
Americans use AI more than ever, and they trust it less. In a 2026 Pew Research Center survey, only 16% of U.S. adults said AI will have a positive impact on society over the next 20 years, while 40% expected a negative one. Bentley-Gallup found the share who believe AI does more harm than good rose from 31% in 2025 to 39% in 2026.

Some of that doubt has been earned in production.
In late October, an AI hacking agent being tested by a security researcher escaped its sandbox, found an internal WorkOS sales tool exposed to the internet, logged in with leftover development credentials and pulled customer records for five days before anyone noticed.
In April, a Cursor coding agent running Anthropic’s Claude Opus 4.6 deleted car-rental software maker PocketOS’s production database and its backups in nine seconds, after finding an overly broad API token in an unrelated file. The outage lasted more than 30 hours.
Even the disputed cases point the same way. The Financial Times reported that Amazon’s Kiro coding agent decided to delete and recreate an AWS environment, causing a 13-hour interruption. Amazon blamed misconfigured access controls, not the AI.
That is exactly the point. When AI fails, the model is rarely the whole story. The failure lives in the access, the trust and the plumbing around it.
That plumbing is what Security Point Break’s new AI Breakpoint Report measures, and I can’t think of a better moment to introduce it.
AI is the Business Process Now
AI is no longer a pilot project off to the side. McKinsey found 88% of organizations use AI regularly in at least one business function, up from 78% a year earlier, and 62% are at least experimenting with AI agents. Gartner projects 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. By 2028, it expects at least 15% of day-to-day work decisions to be made autonomously by agentic AI.
The agents are already inside.
BeyondTrust’s Phantom Labs found a 466.7% year-over-year increase in AI agents across the enterprise environments it monitors. In a 2026 survey by Saviynt and Cybersecurity Insiders, 71% of CISOs said AI tools already access core systems such as Salesforce and SAP.
Check Point says the average organization now runs ten AI applications a month. And the agent is about to stop being an app you choose to install. According to Check Point, Microsoft plans to build AI agents directly into Windows by the end of 2026.
Visibility Hasn’t Kept Up
Yet only 16% of those CISOs said they govern that access effectively. Seventy-five percent have found unsanctioned AI tools running in production, and 95% doubt they could detect misuse. In Pentera’s survey of 300 U.S. enterprises, only 1% reported a dedicated AI security budget.

The tools themselves are exposed. When Check Point reviewed 10,000 Model Context Protocol (MCP) servers, the connectors that give agents their reach, 40% carried security weaknesses. Cobalt’s pentesters find high-risk flaws in AI and LLM applications at 2.7 times the rate of traditional software, and only 38% of those high-risk findings get fixed.
The Breakpoint Count
SPB’s own count tells the same story. From January through October 2026, the AI Breakpoint Report counted 1,507 severe vulnerabilities, rated CVSS 8.0 or higher, in software whose main job is AI. They span 419 products from 361 makers, and 582 of them (39%) are critical at 9.0 or above.

- The pace is accelerating. The monthly count went from 50 in January to 244 in September, nearly fivefold.
- Most are easy to reach. About 62% can be exploited over a network with no login at all.
- Agents take the brunt. More than half, 54%, hit software that builds AI agents or connects them to tools.
- The bugs are old. 65% fall into classes the industry learned to prevent decades ago: injection, path traversal, missing authentication.


October alone brought 220 severe flaws, 75 of them critical and seven of them perfect 10s. Read the October 2026 AI Breakpoint Report.
New models arrive faster than anyone can count them. The AI space is moving faster than regulation, oversight and maybe common sense. We would be remiss to study only how attackers use AI. That is why SPB is focusing on AI as the attack surface, not as the newest adversarial weapon.
The Target, Not the Weapon
This report does not track how attackers use AI to exploit existing flaws, and it does not map AI-assisted attack chains. It tracks flaws in the AI tools themselves.
Those are different problems. In September 2025, Anthropic detected an espionage campaign it attributed with high confidence to a Chinese state-sponsored group it calls GTG-1002. The attackers turned Claude Code into an autonomous operator that handled 80% to 90% of the tactical work, from reconnaissance to exfiltration. That is AI as the weapon. It’s an important story, but it isn’t this report’s story.
Compare Langflow, IBM’s open-source platform for building AI agents. In March, CISA added CVE-2026-33017, an unauthenticated remote code execution flaw rated 9.3, to its Known Exploited Vulnerabilities catalog. Anyone who could reach an exposed Langflow server could run code on it. No AI misbehaved and no prompt was involved. It was an AI tool, deployed in a hurry, with a door left open. That is AI as the target, and it’s what this report counts. All eight 2026 AI flaws on CISA’s exploited list sit in just three products: Langflow, LiteLLM and MLflow.
We also look past the obvious providers. The frontier labs, including Anthropic, OpenAI and Google, are part of the picture. So are self-hosted open-weight models such as Qwen and Kimi, and the sprawling ecosystem of frameworks, gateways and connectors built around all of them. That ecosystem is where the market is headed, and right now it’s where the risk is concentrated.
Find Your Stack
This isn’t a report about AI security tools. It’s about AI as a tool, and AI now runs through the enterprise at every level: customer-service chatbots, coding assistants, the shadow AI tools employees adopt without asking and, increasingly, the security stack itself.
So every flaw is sorted by what the software does, not only by who made it. Start with the categories your organization runs (full breakdowns are in the October 2026 edition):
AI Breakpoint · Find your stack
Start with the AI your organization runs
Each row jumps to that category in the October 2026 edition. Bars show the category’s share of all 1,507 severe AI flaws counted so far in 2026.
Build or run AI agents Agent platforms 80flawsJump → Buy chatbots, assistants or AI plugins AI apps and chatbots 35flawsJump → Connect agents to tools and data MCP connectors 22flawsJump → Train, track or deploy models AI/ML build tools · Model management 19flaws (7 + 12)Jump → Host models themselves Model-serving software · plus model operations, under “Other” 43flaws · a 2026 highJump → Route model traffic through a proxy AI gateways and controls 11flawsJump → Ground models in company data RAG and vector infrastructure · under “Other” 5flawsJump →Rows are ordered by 2026 share. Shares are rounded. “Other” in the October edition combines RAG and vector databases (5) and model-operations tools (5). See how every category moved from September to October. Source: SPB AI Breakpoint Data.
About the AI Breakpoint Report
The AI Breakpoint Report is a monthly count of severe vulnerabilities in AI software: the platforms, models, agents, connectors and applications organizations now rely on for operations, research, security and critical infrastructure.
It fills a gap. Vendor reports describe threats through the lens of the products they sell. Patch Tuesday roundups cover one company’s fixes at a time. Practitioners, CISOs and risk owners have had no neutral, cross-vendor count of what is breaking across the AI stack. This is that count: bugs, patches and in-the-wild exploitation, without the spin.
Every flaw we count has already been published. What we add is context: month-over-month comparisons, year-to-date trends and the patterns that only show up when you look at the whole AI stack at once. When our numbers are uncertain, we’ll say so.
Why October Arrives in November
Each edition covers one calendar month and publishes about two weeks after that month ends. The delay is deliberate. A vulnerability’s first public record is rarely its final form: severity scores are often added or revised days or weeks later, advisories get reviewed and corrected, patches ship after disclosure, and CISA can flag a flaw as exploited at any time. Counting on the first of the month would undercount the month and understate the patches. Waiting lets most records settle before we count them.
Some won’t settle in time. Records that arrive after an edition’s cutoff are added to the month they belong to in our year-to-date totals, so earlier monthly figures may rise slightly over time. Each edition states its data cutoff. Published editions aren’t rewritten; corrections are noted on the edition they affect.
How We Count
We count published vulnerabilities rated CVSS 8.0 or higher in software whose main purpose is AI. At 8.0 and above, a flaw is generally reachable over a network and can do broad damage, which is the profile attackers exploit at scale. Each flaw is labeled AI-stack, a classic bug in AI software, or AI-native, an attack that arrives through content the model reads, such as prompt injection.
An AI model does the first pass at classifying each advisory under fixed written rules. When it isn’t confident, it doesn’t guess: it marks the record Needs ruling, and an SPB editor decides. Before each edition, an editor also reviews borderline records by hand, and every ruling is logged. The full rules, sources and known limits are in our methodology.
Corrections: if you think we got something wrong, email contact@securitypointbreak.com. We’ll fix it and note the change.
Sources: FAA, Nationwide Ground Stop Q&A · TechRepublic on FAA SMART · BleepingComputer on the Microsoft 365 outage · Gallup on local data center opposition · Pew Research Center survey, as reported by Gizmodo · Bentley-Gallup · WorkOS security bulletin · BleepingComputer on WorkOS · IANS on PocketOS · The Decoder on the AWS Kiro outage · McKinsey, The state of AI in 2025 · Gartner on agentic AI by 2028 · UC Today on Gartner’s 40% forecast · BeyondTrust Phantom Labs, via iTWire · Saviynt, 2026 CISO AI Risk Report · Pentera, AI Security & Exposure Benchmark 2026 · Cobalt, 2026 State of Pentesting Report · Check Point Research, AI Security Report 2026 · Check Point, Cyber Security Report 2026 · Anthropic, GTG-1002 report · NVD, CVE-2026-33017 · CISA Known Exploited Vulnerabilities catalog