AI × Application Security × Insurance

AI-written code is outrunning insurance security teams.

Developers now report that about 42% of the code they commit is written or assisted by AI, most of what today's models produce ships with an exploitable flaw, and the time to fix one before it's exploited has collapsed toward zero. Regulators and cyber underwriters moved on it before most security teams did.

Book a working demo Read the teardown ↓

Four forces landed on insurance security in one planning cycle

3 days

federal deadline for the highest-risk fixes, down from 14-15

CISA BOD 26-04

~50%

of AI-generated code that actually works is still exploitable

BaxBench, ETH Zurich

<1 day

disclosure to exploitation, down from 2.3 years in 2019

CSA

10×

more security findings in six months, same repos

Apiiro

Why this stopped being abstract

The code changed, and the clock did too

The code

Veracode's March 2026 re-run found 45% of AI-generated samples still ship an OWASP Top-10 flaw, flat even as the models get better at writing code that works. BaxBench, which writes real exploits against generated backends, is starker still: roughly half of the code that passed its functional tests was still exploitable.

~50%

of functionally correct AI code is still exploitable

19%

of organizations have full visibility into where AI is used

The clock

Regulators put the fix on the security leader's signature. Under NYDFS Part 500 a carrier's highest-ranking executive and CISO personally certify material compliance every year, and Part 500 requires timely, risk-prioritized remediation. EU DORA escalates missed patch deadlines under national penalty regimes reaching 5-10% of turnover. CISA BOD 26-04 cut the highest-risk clock to three calendar days.

So whatWhen you and your CEO sign that you remediate "promptly, based on risk," is that a statement your remediation data could defend, or one you hope nobody checks?

The regulatory clock · 2023 → 2027

"We are working on it" stopped being an answer.

Dec 2023passed

NAIC Model Bulletin on AI; named senior accountability, now live in about half of US states.

Oct 2024passed

NYDFS AI cyber guidance atop timely, risk-based remediation and dual CEO/CISO certification.

Jan 2025in force

EU DORA — weekly scans, patch deadlines, escalation; national penalties reaching 5-10% of turnover.

Dec 2025passed

Colorado second annual ECDIS attestation for life insurers; governance framework (Reg. 10-1-1) now extends to auto and health.

Jun 2026issued

CISA BOD 26-04 — highest-risk fixes cut to 3 calendar days (from 14-15), with forensic triage. Federal, but the benchmark everyone now cites.

End 2027delayed

EU AI Act high-risk obligations for life/health pricing & risk assessment. High-risk deadline deferred to Dec 2027 by the Digital Omnibus, adopted June 2026.

The regimes do not overlap cleanly, and no carrier has been publicly sanctioned under the AI-specific provisions yet. That makes the first enforcement action the reference case for everyone else.

The rules tightened on one clock. The attackers were already running a faster one.

Zero Day Clock · disclosure → exploitation

The time to fix a flaw before it's used has collapsed.

The Zero Day Clock (Sergej Epp, cited in CSA's Mythos paper) puts the gap between disclosure and exploitation at about 2.3 years in 2019 and under a single day in 2026. CrowdStrike found 42% of exploited vulnerabilities in 2025 were hit before public disclosure. The 2026 bar is drawn at a minimum visible height to stay legible beside 2019, not to true log scale. The Zero Day Clock is a leading indicator, not a measure of realized damage; cited as such.

Apiiro · production repositories · Dec 2024 → Jun 2025

AI didn't just add bugs. It shifted the mix toward the dangerous ones.

Same six months, same repos: 3–4× faster commits, ~10× more security findings. The shallow classes fell, while the classes that take real architectural work to fix climbed. All four bars share one scale. Source: Apiiro, Sept 2025, re-confirmed in the CSA "Vibe Coding's Security Debt" note, Apr 2026.

Another scanner won't help. Only 18% of findings rated "critical" stay critical once runtime context is applied (Datadog 2026), and the average organization absorbed roughly 865,000 alerts in a single quarter. Most carriers surface far more genuine, ranked risk than they can act on.

Detection isn't the bottleneck. Resolution is.

What actually moves the number

Two motions, one context graph

If the constraint is fixing, that is the part to automate. Pixee runs two motions on one context graph, on the stack you already own.

Reactive · Triage & Fix

Clear the noise, then ship the fix.

It reads your existing scanners' output and clears the noise, so your team sees the findings that are actually exploitable in your code instead of the whole pile — up to 95% fewer false positives. For the real ones, it writes the fix as a reviewed pull request in your own coding style, not a generic AI patch. Your developers approve or reject it like any other PR, and they merge about three in four (a 78% merge rate, measured across production customer deployments). Deployed self-hosted it runs in your own environment, so source code never leaves, and it keeps a record of every decision — exactly what an examiner or underwriter asks to see.

Proactive · Foresight

Prevent the next one at the spec.

Automated remediation is strongest on the shallow flaws AI is already reducing, and weakest on the privilege-escalation and design flaws it is multiplying. So Foresight moves earlier. It reviews the design before an agent writes the code, then checks that the shipped code kept the plan. As coding agents write more of the software, the specification becomes where security is won or lost. The cheapest vulnerability is the one you never introduce.

one context graph Every fix teaches the next design review, and every design review shrinks tomorrow's backlog. The number comes down from both ends.

Together they move remediation from the financial sector's 276 days to fix half its flaws toward days, with a human reviewing every change before it merges.

Triage & Fix + Foresight · one platform, GA

Trusted by teams at

Your peer set, from the data: we run AppSecJobs.com, an AppSec-only board that discovers roles across roughly 120 employer boards. Carriers are now staffing dedicated AI-security functions: one top-10 auto insurer has posted repeat Senior Staff Engineer, AI Security openings since late 2025. Judging by what carriers are staffing, the work they want to speed up is the deciding and the fixing, not the finding. Directional read of our own hiring corpus, not a survey.

Where a program actually stands

01Can your remediation data back your next attestation? Pull the real time-to-remediate distribution for your open KEV-listed criticals and set it beside the sentence a named executive is about to sign.
02Did review scale with AI-written code? Get the share of code merged last quarter that was AI-assisted, then check whether review coverage grew with it, not headcount.
03What happens to a finding that misses SLA? A documented exception and risk-acceptance workflow is the artifact auditors ask for before the SLA number itself.

See it on your own repositories

Book a working demo

No rip-and-replace. Pixee runs on the 12 scanners you already have natively, plus anything that speaks SARIF, self-hosted in your own environment.

Read the full teardown

AI-Written Code Is Outrunning Insurance Security Teams: An Industry Teardown. The primary sources behind every number above, plus a reading list for insurance security leaders. No form.

Download the PDF ↓

Book a demo