Cloud security provider Wiz’s recently launched Scan for Good initiative has uncovered more than 475 high-severity and critical exposures across public sector, healthcare and open-source environments.
As offensive AI capabilities mature, the cybersecurity community faces a growing challenge. While malicious actors increasingly use AI agents to discover zero-days and probe perimeter defenses, under-resourced public sector entities, healthcare providers and open source maintainers lack the capacity to keep pace.
To help tilt the scales back toward defenders, cloud security provider Wiz launched Scan for Good in September. The initiative pairs autonomous AI scanning with human security researchers to uncover public exposures and complex attack paths across critical infrastructure at no charge.
The service was developed by Wiz in partnership with Google DeepMind and the U.S. Cybersecurity and Infrastructure Security Agency. Wiz played a leading role in uncovering the high-profile PHP2Shell memory safety vulnerabilities that surfaced last month.
According to Wiz, the Scan for Good program has monitored more than 326,000 public endpoints across 17,700 domains since launch, uncovering more than 475 high and critical exposures.
“Our selection process prioritizes organizations where a successful attack could cause meaningful harm to communities and vital societal functions,” Gal Nagli, head of offensive AI security at Wiz, told TechTarget Cybersecurity.
“Eligible entities include registered nonprofits, open source maintainers, educational institutions and public health agencies that lack dedicated in-house offensive security teams,” Nagli said.
Uncovering multistep attack chains
Unlike large enterprises with dedicated security operations centers, under-resourced public and open source environments often lack visibility across their application footprints. This leaves basic configurations unmonitored over time.
Rather than discovering isolated software bugs, Wiz’s initial findings show that exposures frequently emerge from multistep attack chains. Among the 475 high and critical discoveries were exposed administrator credentials on nationally significant archives, unprotected mobile alert channels in public hospitals and leaked database sessions in public transportation networks.
“What we’ve seen is that, rather than single code-level flaws, findings frequently consist of multistep attack chains combining missing access controls, exposed keys/tokens, unsafe uploads, NoSQL injections and public CI/CD credential leaks,” Nagli said.
Driving this discovery service is specialized AI tooling within the Wiz Red Agent, powered by Google DeepMind frontier models, including Gemini 3.8 Flash Cyber and Gemini 4 Argon. Rather than relying purely on traditional static or dynamic code scanners, these models provide advanced reasoning capabilities that map complex relationships across identities, permissions, APIs and application behaviors.
“These AI models have cyber capabilities that help us defend at the frontier and uncover vulnerabilities that would otherwise remain undetected,” Nagli said. “Using the frontier reasoning of these models that are specifically trained to become better patchers and fixers helps teams fix faster, which is yet another area where AI tips the scales towards defenders.”
Validation vs. authority
While autonomous AI agents can map perimeter exposures at machine speed, technical validation represents only the first step in a broader security lifecycle.
As Neil Osnato, founder of Persistence Analytics Group, noted, proving a vulnerability exists is fundamentally different from having the operational mandate to fix it.
“Technical validation does not by itself determine institutional authority to act,” he told TechTarget. “A security team may have enough evidence to conclude that an exposure is technically real without automatically having enough evidence, or authority, for every downstream action that could follow from that finding.”
For example, Osnato explained, an exposed administrator credential or multihop attack path might be technically validated. But deciding to revoke that credential, alter production routing or isolate a hospital system involves operational consequences that require explicit human leadership, materiality assessments and clear governance frameworks, he said.
“The more autonomous these systems become, the more important that distinction becomes,” the analyst said.
By combining autonomous AI discovery with rigorous human validation, Scan for Good could provide a blueprint for how security teams can use offensive AI defensively. This would ensure advanced technology serves the organizations that protect communities before adversaries can exploit them.
“The first priority should be what is truly exploitable today,” Nagli said. “What damage could an attacker cause, and how do you prevent that?”
He added: “We help our customers answer these questions by pinpointing where teams should focus their efforts. That objective translates seamlessly to Scan for Good, where our goal is to surface public exposure and a list of recommendations so that participating organizations can be proactive, before an adversary exploits these vulnerabilities.”
James Walker is lead editor at TechTarget Cybersecurity.
