Microsoft shows editing fewer than 10 neurons can plant a hidden backdoor in AI agents
Microsoft researchers described a technique called Lazarus that changes a few dormant neurons in an open-weight model so that a trigger makes it output attacker commands, with no retraining needed. The modified model behaves normally in standard safety tests, and Microsoft says it has not seen attackers use the method.
Teams running agents on open-weight models need to verify where their model files come from and limit what agents can execute.
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
-
Railway breach let hackers run up about $740,000 in AI charges with stolen customer keys
Platform provider Railway said AI provider keys stored by customers were stolen in an October breach and used for an estimated $740,000 in unauthorized usage, potentially affecting up to 100,000 workspaces. The company promised to cover charges that AI providers will not refund.
Why it mattersStored AI API keys are now a direct target for theft and resale of compute, so set spending limits and rotate keys kept on third-party platforms.
-
Claude Code settings files leaked over 31,000 secrets on public GitHub repos
GitGuardian found 31,088 distinct secrets, including database credentials and cloud tokens, in Claude Code local settings files across 23,276 public GitHub repositories. Of those, 5,343 were still valid when detected.
Why it mattersIf you use Claude Code, check that your local settings file is excluded from Git and rotate any credentials that may have been committed.
-
Google releases Gemini 4 Carbon, its strongest model yet, with an agent 'swarm' feature
Google is rolling out Gemini 4 Carbon to the general public after first giving it to a select group of cybersecurity defenders. Google says it pushes beyond its Argon model, though Argon remains the suggested choice for most coding and knowledge work.
Why it mattersDevelopers and users now have a new top-tier Google model to compare against the best systems from Anthropic and OpenAI.
-
Anthropic releases Claude Fable 5.2 and Mythos 5.2 and cuts API prices by 20%
Anthropic introduced Claude Fable 5.2 and Mythos 5.2, which share one underlying model with different safeguards and bring gains in coding, reasoning and vision. Fable is broadly available, while Mythos is limited to vetted cybersecurity and life sciences groups, and both now cost $8 per million input tokens and $40 per million output tokens.
Why it mattersDevelopers get a stronger coding model at a lower price, and the split release shows how labs are gating the most capable models for sensitive uses.
-
DeepMind's AlphaProof Code formally verifies Signal's cryptographic core in four weeks
Google DeepMind's AlphaProof Code mathematically verified the core cryptographic code of Signal in four weeks, a task that usually takes years, with AI agents writing more than six million lines of proof code. DeepMind plans to apply the method to critical infrastructure such as power grids and medical devices.
Why it mattersAI could make formal proof of software correctness practical for security-critical systems that people rely on every day.
