Google DeepMind and Wiz have launched Scan for Good, a programme using AI with human security researchers to find and help fix exposed systems across public services, critical infrastructure and non-profits. Its early case studies include a hospital alert system, a rail operator, a national archive and a municipal data service. The important point is not simply that AI can find vulnerabilities. It is that the programme is aimed at organisations that may lack the people and time to hunt for attack paths themselves.
Google DeepMind Watch analysis
What happened
Wiz says Scan for Good will use its Red Agent and other internal AI tools to examine publicly facing websites, APIs and applications, but only where testing is authorised through a bug bounty, disclosure policy or explicit permission. Google DeepMind is contributing Gemini models, including Gemini 3.8 Flash Cyber, while Wiz researchers validate findings, contact affected organisations and support remediation.
The company says its systems have already uncovered hundreds of public exposures that were fixed. Its examples include an exposed administrator key with access to 8.8 million archive files, a hospital booking flaw that exposed patient identifiers and clinical information, and a leaked database that could have allowed control of a rail operator’s routes and announcements. Wiz says those issues were privately disclosed and remediated, although the organisations are not named.
Key findings
- Public-sector exposure
Missing access controls and exposed credentials created routes to sensitive data or administrative control. - Human validation
AI can investigate and explore systems, but Wiz says researchers remain responsible for confirming impact and disclosure decisions. - Authorised testing
The programme says it will not scan organisations without permission, a modest but essential detail in a field where enthusiasm can otherwise become trespassing. - Defensive focus
The aim is to find weaknesses before attackers do, particularly in under-resourced services and critical infrastructure.
Why it matters
Security teams have long faced a scale problem: internet-facing systems are numerous, configurations change constantly and a small permissions error can become a much larger attack path. Wiz says AI can connect clues across identities, APIs, permissions and application behaviour faster than a manual process usually allows.
That does not make the system an autonomous security oracle. The reported results come from Wiz, and the examples are company-selected case studies rather than an independent measurement of Gemini’s performance. The safeguards matter because automated probing of public systems can itself create risk if authorisation, validation and disclosure are treated as afterthoughts.
Our read
This is a substantial real-world test of the more useful version of offensive AI: not a demo that finds a clever bug, but a process that helps close ordinary, consequential weaknesses in places people rely on. The interesting question is whether the programme can scale its defensive reach without scaling false positives, accidental disruption or a new industry habit of scanning first and asking permission later.
For organisations responsible for public-facing services, the practical takeaway is straightforward: review your authorisation and disclosure routes, check exposed credentials and permissions, and assume that attackers are improving at joining seemingly minor weaknesses into a workable path.
What to watch
- Whether Scan for Good publishes anonymised technical research after remediation.
- How many assessments are completed, and how many findings lead to fixes.
- Whether independent organisations can reproduce the claimed advantages over conventional scanning.
- How the programme handles false positives, sensitive systems and testing boundaries.
Discussion spark: Should AI-assisted security testing become a standard public-interest service, or does its usefulness depend on safeguards that most under-resourced organisations cannot realistically police?
Sources and evidence
not affiliated with, endorsed by, or operated by Google or Google DeepMind