99.9% of Fixable AI Vulnerabilities Go Unpatched. If You Sell to Enterprise, That's Your Compliance Moat.
A vulnerability intelligence firm (whose data comes from scanning Stripe-verified, real production environments) ran numbers on AI security posture across companies shipping with LLM wrappers, agents, and coding assistants. The findings are bleak:
81.2% of companies running AI packages have at least one known, fixable vulnerability.
Of all vulnerability alerts where a fix is available, 99.9% remain unpatched.
That's not negligence at scale. That's a process failure. And broken processes are where consultants get hired.
Why This Happens
The reason isn't stupidity. It's structural.
A typical company in 2026 has AI in five or six places: Claude wrapped in a Node.js backend (via the SDK). Langflow running agent workflows. GitHub Copilot in their CI/CD pipeline. Maybe a custom fine-tune on Anthropic or OpenAI. Cursor or Windsurf running in local dev environments. A chatbot built on Vercel AI SDK.
Each of these has dependencies. The Node.js SDK has dependencies. Langflow has dependencies (that Langflow 1.8.x has a critical CVE in the dependency tree, CVE-2025-3248, that lets an attacker run code if they can inject into a prompt or config). GitHub Copilot integration has dependencies. The chatbot has dependencies.
When Langflow 1.8.2 ships with a fix for CVE-2025-3248, the company gets an alert. The alert lands in Slack. The security team reads it and thinks: "That's in Langflow. Do we run Langflow in prod?" They ping the infrastructure person. Infrastructure person says "Yeah, on the staging cluster." Staging cluster. Not prod. So it's low priority. It gets added to a backlog. That backlog is two sprints deep. Two sprints later, something more urgent happens. The fix doesn't ship.
The same cycle repeats for Cursor's DuneSlide fixes, for GitHub Copilot prompt injection patches, for symlink escape mitigations in agent code. Each one is low enough priority on its own that it doesn't bubble up. The aggregate (81% of companies with at least one unpatched vulnerability) is the water you're swimming in.
Then the FTC, the SEC, or the customer's compliance team audits. They ask: "Show me your AI vulnerability remediation process." The company pulls up: Jira tickets. A spreadsheet. Maybe a Confluence doc that's three months out of date. The auditor looks at the spreadsheet and sees three pages of "Vulnerability found – no fix" and "Vulnerability found – fix available – status: not started." The company fails the audit or gets cited for missing controls.
That's the gap. And that gap is a project.
What the Compliance Angle Actually Looks Like
Here's what a $10K-$50K retainer looks like:
Month 1: Inventory and baseline assessment. You audit the customer's AI footprint. You find every LLM SDK, every agent framework, every coding assistant integration. You run them against the latest CVE databases (NVD, GitHub Advisory, Trivy, Snyk). You produce a report: "You have 27 AI dependencies. 19 have updates available. 7 have known vulnerabilities. 3 of those 7 are in your production path."
Month 2–3: Remediation plan and automation. You help them build a process. Not "check manually every week": a real process. You set up dependency scanning in their CI/CD (Dependabot, Snyk, Trivy, it doesn't matter which; they all work). You create an automated monthly report. You define triage criteria: "Critical CVE in prod path = patch within 72 hours. High CVE in staging = patch within two weeks." You get someone in the room to sign off on those thresholds.
Ongoing: Monthly audit and escalation. Every month, you run the scan again. You produce a report showing what's been fixed, what's new, what's overdue. You escalate anything that's overdue to the stakeholder who promised to own it. You document the audit trail (this is the compliance win: when the regulator asks "how do you manage vulnerability risk," you show them the dated, signed-off reports showing you're actively tracking and remediating).
The hourly rate depends on your market. For a 50-person startup in a regulated vertical (fintech, healthcare), $150/hour = $2K–$5K per month. For a 500-person mid-market company in insurance, $200–$250/hour = $5K–$15K per month. For an enterprise with complex AI stacks, $300+/hour or a flat monthly retainer = $20K–$50K+.
Why This Works for a Solo Operator
An enterprise security firm can't afford to care about "audit Langflow on your staging cluster for CVEs." Their minimum engagement is six figures. But you can own this for a specific vertical or vertical segment, and the overhead is low:
- You get familiar with the top 20 AI frameworks and their CVE histories. You know Langflow's issues, GitHub Copilot's sandbox escapes, Cursor's RCE variants, Mistral's supply chain risks. You can run an inventory audit in two hours per customer.
- You automate the scanning. Trivy + Dependabot + a script that generates a monthly report = four hours of setup per customer. Then it runs by itself.
- Your compliance knowledge is deep in one vertical. "Here's how a healthcare company should handle AI vulnerability disclosure," or "Here's what your fintech regulator expects to see in audit trails." That depth is worth $150–$300/hour.
The enterprise security firm has a process that works for $500K engagements. You build a process that works for $5K–$15K engagements. Different customers. Different leverage.
The Honest Take
This is unsexy work. Vulnerability management is boring. Scanning is boring. Writing reports is boring. The customers who buy it often don't celebrate it (because the job is "nothing bad happens," which feels like nothing).
But that's exactly why it's defensible. The shiny "we'll build you an AI agent that does X" work commoditizes fast. Everyone can build an agent. The boring "we'll make sure your AI stack doesn't get exploited" work is hard to commoditize because it requires context, continuity, and trust. Once a customer knows you're running their monthly scans, you're sticky. You're the person they call when a zero-day drops on a Friday.
And the regulatory environment is hardening. The SEC is starting to ask companies for AI governance frameworks. The FTC has guidance on algorithmic accountability. By 2027, any company shipping AI features to the public is going to need to demonstrate, in writing, that they have a process for vulnerability remediation. You're not selling a luxury. You're selling a requirement.
What I'd Actually Do
If you're deep in one vertical (healthcare, fintech, insurance, legal), start with the three largest companies in your network. Offer to do a one-time AI vulnerability audit for free or at a reduced rate. The audit takes four hours. You produce a 10-page report showing what you found and what a monthly remediation service would cost.
If any of them bite, you have a template. You can replicate it for the next five companies in that vertical. After three or four retainers, you have enough data to build a light SaaS layer on top (a dashboard, automated reporting, client portal). At that point, you can raise rates or hire someone part-time to help with scans.
Don't try to be Snyk or Trivy. You're not building a platform. You're building a service that speaks the language of "here's what your regulator is going to ask about and here's how we ensure you're compliant." Sell the compliance story. Use the existing tools. Pocket the margin.
Author
Lukas
@lukcombinator