Cyber insurers are pricing based on patching speed

Your cyber insurer has a new question, and it is not whether you found the vulnerability. It is how fast you fixed it. As AI reshapes both sides of security, underwriters and boards have started to grade security programs on remediation speed, because that is where a loss now happens, or doesn't. As a result, cyber insurers are pricing remediation speed into how they underwrite policies.

The exploit gap

The numbers driving this shift are stark. According to Verizon's 2026 Data Breach Investigations Report, exploiting a known vulnerability has become the most common way for attackers to break in, accounting for 31% of breaches, up from 20% the previous year.

At the exact same time, enterprise remediation efforts are moving backward:

• The Defender's Clock: The median time to fully resolve a critical vulnerability has risen to 43 days. Furthermore, only 26% of the vulnerabilities in CISA's Known Exploited catalog were fully remediated, down from 38%.
• The Attacker's Clock: At Endor Labs, we have tracked new CVEs going from initial disclosure to active exploitation in under 10 hours. Worse, Mandiant's M-Trends 2026 now puts the mean time-to-exploit at negative seven days. Attackers are weaponizing flaws before the official patch even ships.

Detection got fast. Fixing didn't.

Detection has never been faster. Frontier models, researchers, and scanners now surface vulnerabilities at a rate that would have looked absurd a few years ago. The work on the other side of the finding still moves at human speed: confirm it is real, find an upgrade that does not break the build, open the pull request, check it did not break three other things. The backlog is what that mismatch looks like when it piles up.

The only way to close the gap is to move the fix onto the same clock as the find, which for most organizations today means handing remediation to agents. Teams that do this report 6x faster CVE remediation and 83% fewer blocked pull requests, and detection-to-production in four to six hours is achievable with current tooling. That last part matters: this is an engineering problem, not a research problem.

But what about token costs?

Point a general-purpose AI coding agent at your findings and you will feel the appeal within an afternoon. You will also meet the reason finance has started paying attention.

Agentic workloads are not chat. A chat turn is one round trip. An agent loops, re-sending its accumulated context at every reasoning step, which is why agent workloads burn 5 to 30 times the tokens of a single chat interaction. This is now enough of a problem that the FinOps Foundation reports 98% of organizations are actively managing AI spend, up from 63% a year earlier, and more than half admit they do not understand the full scope of what they are spending.

The costs are real and they scale with every pull request. One 50-engineer team ran a per-PR security agent and watched it generate an $8,400 bill in the first month, the root cause being a 50,000-token security policy injected into every single prompt. Building native AI-powered PR and code security for an enterprise with 150 developers pencils out to roughly $900,000 a year at current frontier pricing, and closer to $1.6 million at the rates of newer models like Claude Fable 5. 

Spending more does not buy correctness on its own: an academic baseline that asked a model only to "find the vulnerabilities" produced nearly 14,000 candidate sinks in a single project and confirmed exactly zero real ones. Expensive and wrong is the worst square on the board.

None of this means the agents are the problem. It means pointing them at everything, blind, is a budget decision no one actually made on purpose.

The agent harness is the answer

We ran the test directly. Take 442 realistic application security prompts across 12 large open-source projects in six languages, covering five tasks: reachability, exploitability, prioritization, remediation, and license compliance. Run them twice, same model, same questions. One agent gets compact evidence from Endor Labs, findings with reachability and call paths, upgrade impact, license risk. The other gets nothing but the repository and its own reasoning.

The agent with context used 6.6 million tokens and finished in 133 minutes. The agent reasoning alone used 79.5 million tokens and took 370 minutes. That is 92% fewer tokens and 2.8 times faster, from the same model, doing the same work.

Measure cost per fixed vulnerability

If speed is what the market now prices, the scorecard has to change with it. Open CVE count gives way to reachable, exploitable, unfixed CVEs. SLA compliance in months gives way to MTTR in hours. Tickets filed gives way to fixes shipped as pull requests.

This is exactly the world we’ve been building for at Endor Labs, and why we shipped AURI Agents a few weeks ago. The are pre-built, context-grounded security agents that run inside the tools your developers already use, or on your own infrastructure with your own model keys. They can autonomously burn down vulnerabilities across your first-party code or open-source dependencies.

A staff security engineer at a global digital ad platform put the payoff plainly: 

The AI SAST triage agent uses dataflow and context to prioritize and patch, exceptions get approved in the workflow instead of a ticket, and security keeps pace with how fast the team ships rather than becoming the bottleneck.

The teams pulling ahead are not the ones spending the most on inference. They are the ones handing their agents the answer instead of asking a frontier model to rediscover the codebase on every loop. That is what turns a 43-day clock into a same-day one, at a token bill you can put in front of your CFO without flinching. 

Book a demo to learn more about using agents for autonomous vulnerability remediation.