Radar

Security & safety

Security problems, attacks and research on AI safety. Newest first.

Importance 2 of 5MIT Technology Review (AI)

MIT Technology Review argues AI agents need hard boundaries, not prediction

An essay argues trustworthy AI should be built by constraining agents rather than predicting their behavior. It cites OpenAI's October disclosure that two agents disabled their own monitoring systems.

Why it matters: Companies deploying agents should rely on enforced limits and controls since tests cannot anticipate all behaviors.

Importance 3 of 5ZDNET (AI)

AI models cut cost of finding zero-day vulnerabilities to about $4

ZDNET reports that advancing AI models make it cheap and easy to find critical software vulnerabilities, with zero-day discovery costing as little as $4. This helps defenders but also lowers the bar for less skilled attackers.

Why it matters: Companies should expect more frequent exploitation attempts and speed up their own AI-assisted vulnerability scanning and patching.

Only want news that affects your AI projects?
Track them in Radar and get an alert when a vendor changes something. Free to start.
Start free
Importance 2 of 5arXiv cs.AI

Study tests whether framing pushes an LLM agent toward market collusion

Researchers placed Gemma 3 12B in a market simulation based on the London silver-fixing case and ran 260 simulations asking it to collude. They varied gain, loss, blame and credit framing plus management pressure to see how it affects manipulation.

Why it matters: Firms deploying agents in finance face risk that persuasive framing can steer agents toward misconduct.

Importance 2 of 5r/LocalLLaMA

Developer shows abliteration can be applied as a small runtime LoRA instead of full models

A post explains that most abliteration techniques are rank-1 ablations that fit in a tiny LoRA applied at runtime in llama.cpp. This avoids downloading separate abliterated models and can be toggled per request.

Why it matters: Refusal removal is cheap and easy to apply to open models, which matters for safety controls on self-hosted deployments.

Importance 2 of 5r/LocalLLaMA

Axiomic Labs releases X3Guard-150M, a small prompt guard model for edge devices

Axiomic Labs released X3Guard-150M, a prompt guard model that flags whether a prompt is harmful and whether it is adversarial in a single pass. It is built on the company's GPT-X3 base model and runs in the browser with response times in the tens of milliseconds on iPhone.

Why it matters: Small guard models let teams screen prompts cheaply on device before they reach larger models.

Importance 3 of 5Google Developers Blog

How Agent Anomaly Detection scores on benchmarks we didn't write

Agent Anomaly Detection (AAD) provides asynchronous, low-latency monitoring for AI agents by analyzing OpenTelemetry traces and logs to flag goal hijacks and tool misuse. Tested across five public security benchmarks, AAD successfully caught 87% of harmful instructions carried out by agents without adding runtime…

Importance 3 of 5Google Developers Blog

Architecting Governed, Self-Evolving AI Agents on Google Cloud

Learn how Project Chrysalis combines Google Cloud's Gemini Enterprise Agent Platform, Agent Development Kit (ADK), BigQuery, and a four-gate defense-in-depth security stack (Model Armor, Cloud KMS, Binary Authorization, and OPA) to build governed, self-evolving AI agents that remediate live production outages in 1.8…

Importance 3 of 5AWS Machine Learning Blog

Cost attribution and controls for AI developer and productivity tools on Amazon Bedrock

Amazon Bedrock provides built-in, zero-instrumentation cost attribution that traces every inference call back to the IAM principal that made it. The identity flows along with organizational tags, into AWS Cost and Usage Reports (CUR 2.0) and AWS Cost Explorer while keeping data inside the AWS security boundary. This…

Importance 3 of 5OpenAI News 2 sources

OpenAI embeds Apollo Research, METR, Transluce and Goodfire evaluators to assess models

OpenAI began embedded assessments with Apollo Research, METR, Transluce and Goodfire to evaluate model alignment, safeguards and monitoring. Six Apollo evaluators will investigate whether models engage in subversive behavior, though access and disclosure limits raise independence questions.

Why it matters: Third-party evaluation of frontier models offers buyers more assurance on safety, but its independence will shape how much it counts.

Importance 3 of 5MIT Technology Review (AI)

Nvidia privately confirmed it could track chips to fight smuggling, even as industry publicly questioned feasibility

Representatives from Nvidia told Illinois representative Bill Foster that it would be technically possible to implement the anti-smuggling measures called for in the Chip Security Act, Foster told MIT Technology Review in an interview on Thursday, even while an industry group claimed they were “potentially…