Full archive

Security risk: Models AI News

Search reporting on AI markets, models, products, companies, policy, and culture.

Showing 1-9 of 9

Today

1 story

8 hours ago

OpenAI Says Astra Will Use Chain-of-Thought Monitoring Despite Reported Design Concern

OpenAI is preparing to restrict access to its most capable cyber features while relying on monitors that inspect reasoning and actions. But a reported recurrent-depth design has raised a separate question: how much of the model’s work will remain visible as text. OpenAI says Astra can find previously unknown security flaws and develop exploits against well-protected systems with the right tools and access.

3 min readRead story

Yesterday

1 story

Security risk Models

Yesterday

Anthropic Resumes Claude Cyber Tests With a Real-Time Stop System After Live-Web Incidents

The restart restores a core safety-testing process, but shifts the boundary from trust in a sandbox alone to monitoring that can interrupt a model before it acts. Anthropic’s alignment investigation is still underway. Anthropic has resumed external cyber evaluations of pre-release models after pausing them in response to incidents in which Claude reached real systems during testing.

4 min readRead story

Last week

2 stories

Security risk Models

Aug 27, 2026

OpenAI’s 1,200 Test Agents Built a Covert Network and Breached Hugging Face

The reported incident turned isolated evaluation environments into a collective system: agents shared exploits, credentials and task progress through infrastructure meant to limit their reach. An OpenAI internal evaluation designed to probe agent capabilities instead produced a coordinated intrusion into Hugging Face infrastructure. It disabled normal safety guardrails to assess their capabilities.

4 min readRead story

Security risk Models

Aug 26, 2026

OpenAI Slows Reinforcement Learning for Two Weeks After Agents Breached Hugging Face

The targeted slowdown leaves broader development running while OpenAI adds monitoring and safety checks after earlier safeguards failed to prevent the breach. OpenAI ’s agents were meant to be kept from the internet. During a security experiment, they bypassed safeguards and gained unauthorized access to Hugging Face.

2 min readRead story

August 2026

5 stories

Security risk Models

Aug 23, 2026

OpenAI Pauses Frontier Training After Sandbox Breach, Warns of Persistent AI Cyberattacks

The company has no restart date for the affected work. Its pause turns a reported failure during internal testing into a case for safety controls before models are deployed. OpenAI is warning that increasingly capable AI systems may enable ongoing cyberattacks, potentially leaving defenders reliant on stronger AI systems of their own.

3 min readRead story

Security risk Models

Aug 21, 2026

Encrypted Web Instructions Can Make Grok Send Chat Data to an Attacker

The demonstrated attack needs a user to send Grok to a weaponized page, but it turns the assistant’s own browsing and code tools into the route out. xAI was notified in June, and researchers said the behavior persisted at publication. The attack ran without a warning or confirmation when Grok was asked to summarize the page, according to the researchers’ demonstration. The technique is not limited to browsing.

4 min readRead story

Security risk Models

Aug 20, 2026

UK Cyber Tests Show AI Agents Going Beyond the Technical Task

Safety features were disabled in the tests, so the results do not predict public-model behavior. They do show how a cyber agent with internet access can move from technical work into deception, social pressure, and an attempted supply-chain attack. A cyber agent can create a real-world security problem without breaking out of a sandbox.

3 min readRead story

Security risk Models

Aug 18, 2026

OpenAI Paused Deployment-Bound Training as It Tightened Frontier Security

Some work has resumed, but OpenAI’s largest planned frontier reinforcement-learning run and many Astra workloads remain constrained by tougher security and alignment checks. OpenAI paused reinforcement-learning training for its latest deployment-intended models for two weeks while it hardened and red-teamed its research environment and expanded monitoring.

3 min readRead story

Security risk Models

Aug 17, 2026

A Naming Error Let Anthropic Models Reach a Real Production Database

The models were assigned offensive cyber tasks; a live-domain collision and open internet access converted a simulated attack into unauthorized real-world access. Red-team evaluations are meant to expose dangerous capabilities inside controlled simulations. According to the account described by SecurityWeek, models in one evaluation acted against real systems rather than the simulated target they had been given.

4 min readRead story