Search the evidence

Find the signal.

Search titles, impact clusters, countries, organizations and the full text of every analysis.

2 stories found

A small citrus key fits a bounded terminal opening while a larger steel key ring remains outside.
SecurityGlobal+1 clusters01

The login was real. The request wasn't.

A genuine sign-in page can still be part of a misleading request. Microsoft's new research demonstrates how a user-approved device-code authorization can expose capabilities connected to a coding agent. The danger is not that looking at a website automatically takes over a computer. It is that a person can authorize a session they did not intend to create. This was a controlled demonstration, not a reported campaign in the wild. Consequences depend on permissions, organization policies and enabled features; remote session misuse, for example, requires relevant remote access to be active. Those conditions are central to the finding, not fine print. The practical defense is to approve only authorizations you initiated and to limit what the resulting identity can reach. There is a more encouraging security story alongside it. Calif reports that human-directed AI research helped uncover a Java vulnerability, and Oracle's November update lists the corresponding fix. That is a different mechanism: finding a software flaw, rather than misleading someone into delegating access. The discovery has hardware- and application-specific conditions; it does not mean every Java installation is equally exposed. Together, the stories show why “AI helps hackers” is an incomplete account. AI can help defenders inspect software while new agent workflows create authorization risks. Neither cancels the other. Our view is that organizations should measure the defensive gains and the new exposure separately, rather than declare AI inherently safe or inherently insecure. A capable tool needs a carefully bounded identity as well as competent people using it.

3 min
A compact cyber model repeatedly searches branching code paths, locating vulnerabilities behind a controlled access gate.
Technical failuresGlobal+3 clusters02

A lightweight cyber model scales vulnerability discovery—and risk

Google DeepMind says Gemini 3.5 Flash Cyber, a lightweight model tuned to find, validate, and patch software vulnerabilities, can outperform larger systems by searching many code paths repeatedly. In testing on the V8 JavaScript engine, it found 55 unique confirmed issues, including 10 missed by the comparison models. The same model generated a reliable remote-code-execution exploit against a production service, illustrating why Google is initially limiting access to governments and trusted partners through a controlled pilot.

3 min