Technology

When AI Finds Vulnerabilities Faster Than Teams Can Patch

Security engineer reviewing code on a laptop and external monitor
Photo by Jakub Zerdzicki on Pexels, used under the Pexels License

AI can now search for exploitable software flaws at a speed that changes the economics of security work. The important question for most companies is no longer whether an AI model can find a vulnerability. It is whether the organisation can turn a growing stream of findings into verified, prioritised and safely deployed fixes.

On 5 November, OpenAI introduced GPT-6 Cyber, a specialised model for authorised security research. OpenAI classifies it as “Critical” in cybersecurity capability and reports that it scored 48.5% on ExploitGym, compared with 32.1% for its previous cyber model. In a black-box web test, it reportedly recalled 40.3% of vulnerabilities, compared with 32% for the general-purpose GPT-6 Astra. These are vendor results, not a guarantee in any company’s environment, but they point to a real shift: vulnerability discovery is becoming cheaper and more scalable.

The bottleneck moves downstream

A conventional scanner usually produces signatures, severity scores and a backlog. A capable cyber agent can go further: inspect code, reproduce a crash, test exploitability, propose a patch and check whether the fix holds. That compresses several specialist tasks into one workflow.

It can also create a new failure mode. If discovery volume rises while remediation capacity stays fixed, teams get a larger queue, not a safer product. Recent coverage of reported Linux and Firefox findings illustrates the scale that security teams may need to absorb. Engineers may waste time on duplicated or unreachable findings, accept a plausible but incomplete patch, or push a fix that breaks production. The scarce resources become trusted evidence, application context, review attention and safe release capacity.

The market is already moving in that direction. On 9 November, IBM described a workflow that connects risk-based vulnerability identification, exploitability prioritisation, targeted remediation and developer-approved pull requests. The pattern matters more than any individual product: discovery and fixing need to be designed as one controlled system.

Build a remediation pipeline, not an autonomous hacker

For founders, operators and technical leaders, a useful operating model has six stages.

1. Define an executable scope contract

Every run should name the repositories, hosts, environments, credentials, time window and permitted actions. Deny everything else by default. A sentence in a ticket is not enough; enforce the boundary with network policy, sandboxing, read-only credentials and short-lived access.

This is especially important because capability and alignment are different properties. OpenAI reports strong scope adherence, but its system card still recommends isolated environments, explicit limits on targets and credentials, action checks before sensitive operations, and human operators who can stop high-risk workflows.

2. Require an evidence bundle

A finding should not enter the engineering queue as prose alone. Require the affected version, reachable path, reproducible test, expected versus observed behaviour, exploit preconditions, severity rationale and relevant logs. Record which model, prompt, tools and artefacts produced it. Evidence makes the result reviewable and prevents a confident narrative from becoming a production fact.

3. Prioritise by business exposure

CVSS is an input, not the queue. Combine technical severity with internet exposure, data sensitivity, tenant boundaries, available mitigations, observed exploitation and the business criticality of the service. A medium-severity flaw on an exposed identity path may deserve attention before a critical issue in unreachable test code.

4. Generate fixes in isolation

Let the agent produce a minimal patch and tests on an ephemeral branch or disposable environment. Do not give it a direct route to production. Keep dependency changes explicit, block unrelated edits and have a code owner approve the diff. For sensitive systems, separate the agent that discovers the flaw from the process that approves the fix.

5. Validate the repair independently

A passing unit test is not closure. Re-run the original proof, add a regression test, execute relevant integration and security tests, and inspect for variant paths. Then release through the normal staged deployment with observability and a rollback plan. The strongest evidence is that the exploit no longer works and legitimate behaviour still does.

6. Measure the full loop

Track verified findings per engineering hour, duplicate rate, false-positive rate, median time from proof to patch, patch rejection rate, regression escapes and reopened findings. Avoid celebrating raw findings: it rewards noise. The outcome metric is reduced, validated exposure.

A practical 30-day move

Choose one bounded application—not the entire estate—and one vulnerability class. Establish a baseline from the previous month. Run the AI workflow in a sandbox against a known snapshot, require the evidence bundle, and route every proposed patch through existing review and deployment controls. At the end of the pilot, compare verified risk removed, engineering time and escaped regressions with the baseline.

Teams should also rehearse the stop conditions: out-of-scope network access, unexpected credential use, attempts to change unrelated files, ambiguous ownership or a patch that cannot be independently reproduced. A safe pilot proves that the organisation can control the workflow, not just that the model can find bugs.

The Qomra Tech view

Cyber-capable AI will reward teams with mature delivery systems. Asset inventory, testable boundaries, clean ownership, reproducible builds and staged releases now determine how much value a company can safely capture from faster discovery. For Gulf organisations modernising critical services, the near-term advantage is not fully autonomous security. It is a disciplined loop in which AI expands the team’s reach while accountable people retain authority over scope, evidence and release.

What to do next: map one high-value application, define its authorised test boundary, standardise the evidence required for a finding, and measure time to a validated fix. That creates a foundation that will remain useful as models become more capable.

Let's talk

Tell us about your project.

We'll come back within one business day with the right person to talk to.

Trusted by founders across healthcare, hospitality and professional services. London HQ · Bilingual EN/AR delivery · NDA-friendly