Microsoft researchers show AI agents can be backdoored by editing fewer than 10 neurons
Microsoft researchers showed a technique called Lazarus that plants a hidden backdoor in an AI model by changing fewer than 10 rarely used neurons, with no retraining. When a trigger appears, an agent using the model runs an attacker's command, yet it behaves normally during standard safety tests.
Companies that give AI agents access to terminals or systems need to verify where model weights come from and who can change them, because safety tests may not catch this kind of tampering.
aii.news picks and summarizes stories from trusted news sources with the help of AI. Every story links to the original reporting. How we work.
Keep reading
-
Over 35,000 apps built with Grok's coding agent found open online, some leaking passwords
Security firm Claroc found 37,571 apps built with Grok online, and 35,578 of them could be opened without signing in, including 24 built by staff at X and SpaceXSI. About one in ten contained internal business data, and some leaked confidential information and passwords.
Why it mattersAnyone building apps with AI coding tools should check access settings before publishing, since default setups can expose private data to anyone.
-
Hackers stole AI provider keys from Railway and ran up $740,000 in fraudulent usage
Cloud host Railway confirmed a nearly 26-hour fight against attackers who hunted for AI provider credentials, generating about $740,000 in fraudulent AI charges. The stolen keys were sought for resale to services used for model distillation, and up to 100,000 workspace credentials may have been exposed.
Why it mattersAI API keys are now valuable loot, so developers should store them securely, rotate them and set spending limits.
-
OpenAI launches Daybreak security tools that use AI agents to pentest apps and prepare fixes
OpenAI introduced Daybreak Red Team, which uses coordinated AI agents to test web applications for weaknesses and is offered through early access. It also launched Defense Factory, which investigates security findings and prepares patches, linking vulnerability discovery with fixing and retesting.
Why it mattersSecurity teams may soon be able to hand both finding and patching of app flaws to AI agents, which could speed up fixes but needs careful human review.
-
Hackers exploit SGLang flaw to breach NEAR AI inference server and steal agent and user data
An attacker used a remote code execution flaw in the open-source SGLang inference engine to run code on one of NEAR AI's inference servers, taking data from 18 agents and about 100 users. Requests sent in a confidential mode were also potentially visible on that server.
Why it mattersTeams running open-source inference engines should patch quickly, because a single flaw can expose user prompts even in so-called confidential setups.
-
Microsoft AI agents and researchers find 59 Linux kernel bugs, some hidden for over 20 years
A Microsoft FORGE Lab project combining AI agents with human researchers found 59 Linux kernel vulnerabilities, some in production code for more than 20 years. AI sped up discovery, but checking flaws and writing patches still needed a lot of human work.
Why it mattersAI is finding old bugs in core software fast, so teams running Linux should expect more patches and plan for the human effort needed to fix them.
