Deep dive: fixing what scanners find

Who owns the fix: what's real

My view as of October 2026: finding cloud risk is cheap and getting cheaper, ranking it is expected of every tool, and the gap is getting fixes merged by the teams that own each service. The evidence on fix capacity is strong and old. The evidence on how much context helps comes mostly from vendors grading themselves. And AI remediation exists, with no published results yet.

What the evidence says about fixing

• Capacity: the Cyentia and Kenna study (about 300 organizations, published 2019) found a typical organization can close about 1 in 10 of its open vulnerabilities in a month, regardless of size; about half fall further behind, and the median vulnerability took 100 days to fix. It predates CNAPP and covers all infrastructure, so I'd use it for the shape of the problem more than the exact rate.
• Known-exploited bugs: secondary summaries of the 2026 Verizon report say only about 26% of vulnerabilities on CISA's known-exploited list were fully fixed, down from 38%, and the median time to patch rose to 43 days. Wiz, citing an outside tracker, puts the gap from disclosure to exploitation at about 21.5 days.
• Code and alerts: Veracode's 2026 report found 82% of organizations carrying security debt, and Palo Alto's Unit 42 reported the average organization's cloud alerts rose 388% during 2024 (a vendor figure).
• Cheap fixes left undone: the metadata setting most instances could enforce without breaking anything, in Rules and policy, and the 59% of IAM users with keys over a year old.

What context really cuts

Every row is a vendor measuring its own customers with its own checks; nobody runs a neutral census of cloud findings. The direction is consistent across competitors, so I believe context cuts the list a lot, but I wouldn't quote any one percentage to a buyer.

Where the fix gets stuck

• No owner: findings on untagged resources with no link to a repository sit unassigned, and the newest survey I found on who fixes cloud findings is from 2021.
• Split budgets: the CISO buys the tool and engineering pays in time, so the team that owns the fix doesn't own the goal.
• Console fixes and open-ended exceptions: the fastest fix drifts back at the next deployment, and suppressions without expiry make the dashboard look better while the risk stays put.
• Ticket overload: I'd expect a tool that files every finding as a ticket to train engineers to ignore the queue. AWS now creates Jira and ServiceNow tickets from findings, and Wiz assigns owners in one click, so routing is getting easier while deciding what deserves a ticket stays hard.

What AI changes so far

• Remediation agents: Wiz's Green Agent, in public preview since March 26, 2026, investigates the top risks, finds the owner, writes command-line, Terraform or Kubernetes fixes and opens pull requests or hands context to coding agents. It doesn't apply changes by itself. No vendor I found publishes how often its AI fixes get merged or how often they break something.
• Testing: AWS's Security Agent became generally available for on-demand penetration testing on March 31, 2026, with no independent data on its quality yet.
• Analysts: Microsoft's randomized trial with 147 security professionals found Security Copilot made them about 7% more accurate and 14-39% faster on lab tasks, while a benchmark from CrowdStrike and Meta found models got malware analysis right only about 15-34% of the time. IBM's 2026 survey links extensive security AI to cheaper breaches ($4.00 million against $5.93 million), a correlation among breached companies.
• Attackers: Anthropic reported a campaign it calls GTG-1002 in which AI agents performed 80-90% of the tasks against about 30 targets, with few successes (Anthropic's own account).

AI investigation of incidents is covered in Observability: metrics, logs and traces.

Securing AI workloads

AI security posture is the posture, data and secrets checks above pointed at AI resources: model endpoints, notebooks, training data buckets, AI API keys and agent identities. It's sold as a module of the platforms more than as a category of its own, and the clouds bundle it. AWS made AI inventory free in Security Hub Essentials in July 2026, Microsoft's paid posture plan covers Azure OpenAI, Azure Machine Learning and Amazon Bedrock, and Google's Model Armor screens prompts, responses and MCP tool calls on any cloud.

The demand is real even if the numbers are soft: Wiz said in 2025 that 85% of organizations host AI in their cloud (a vendor figure), and IBM's 2026 survey found unsanctioned "shadow AI" in 43% of breached organizations. Runtime filters for prompts and tool calls are usually a separate purchase, and how agents get identities and tool permissions is in AI agents: orchestration platforms.

Questions to ask a vendor about fixes

1. Of the issues you ranked critical for us last quarter, how many did our teams close, and how fast?
2. How do you find the owner of a resource with no tags, and what share of our findings have one?
3. Does the fix go back into code, and do you flag it when a console fix drifts?
4. Do exceptions require an expiry date, and do suppressed findings still feed your threat detection?
5. When your AI opens a pull request, what share are merged unchanged, and who's accountable when one breaks something?