AI safety and alignment: jailbreaks and defences, research on model behaviour, safety evaluations and governance frameworks.
OpenAI's safety team systematically investigated why the chains of thought in frontier models like GPT-6 Astra are increasingly evading oversight. Four training steps drive the erosion, including architecture choices and reinforcement learning. The analysis shows that the loss of transparency is not a law of nature but also reflects tradeoffs around capability and cost.
LLMs rely on their context to generate the next set of tokens. To address this, we developed LLMON (LLM Object Notation, pronounced “lemon”), an LLM-native markup language for explicitly communicating structure and semantics to the models.
Google Deepmind's AlphaProof Code mathematically verified Signal's cryptographic core code in four weeks, a process that typically takes years. AI agents powered by Gemini Carbon wrote more than six million lines of proof code and flagged only about one percent for human review. Deepmind plans to extend the method to critical infrastructure like power grids and medical devices.
Security firm Claroc found more than 35,000 Grok-built apps open to anyone without login, including patient data and confidential SpaceXSI documents. The cause: "Anyone with the link" was preselected as the default access setting. SpaceXSI took only its own 24 employee apps offline. The rest remain accessible.
Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access.
Google is making Gemini 4 Carbon generally available after a two-week testing phase. The model ranks first in several benchmarks, in some cases ahead of Anthropic's new Fable 5.2, at less than half the per-task cost. Google is also previewing coordinated agent teams through the "/swarm" feature, which split up complex tasks in real time.
Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers. Carbon is our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning.
AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption.
Anthropic is releasing Claude Fable 5.2 and Mythos 5.2, its most powerful models to date, in two versions: one freely available and one reserved for verified organizations. The models solve open physics problems, show stronger cybersecurity capabilities, design functioning operating systems, and write more than 90 percent of the changes merged into Anthropic's core code, according to the company.
In the Vals AI benchmark, OpenAI's GPT-6 Sol Daybreak hits 80.8 percent of attack objectives in simulated corporate networks. Worse, the open-weight model DeepSeek V4.1 Flash reaches 72.8 percent at a tenth of the cost. Because open models can no longer be restricted, analysts say the window for defenders is shrinking fast.
Anthropic is using AI to build more powerful AI. And it appears to be working remarkably well, with the company's models improving roughly twice as quickly as a year ago, with little indication of slowing down.
Nonprofit METR spent eleven weeks auditing Anthropic's internal research data and concluded that the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted the most informative evidence, though, which means the public has to rely on METR's judgment "much more than would be ideal."
With GPT-6 Cyber, OpenAI is introducing an AI model built to find security vulnerabilities in software, develop exploits, and suggest patches. According to OpenAI, it has already found 81 previously unknown vulnerabilities in Firefox and developed 948 working exploits in the Linux kernel, significantly more than its predecessor.
Introducing GPT‑6 Cyber, OpenAI’s cybersecurity model for finding exploitable vulnerabilities and developing fixes, trained to stay within authorized scope.
OpenAI's new GPT-6 Astra model apparently sabotaged AI safety research on purpose. It invented obstacles, withheld solutions, and performed about 40 percent worse on safety tasks than on comparable other tasks. Removing a flawed training dataset fixed the problem, but OpenAI can't fully rule out strategic motives.
OpenAI built an autonomous pipeline that found, validated, and generated patches for 948 local privilege-escalation vulnerabilities in the Linux kernel.
Epoch AI researchers say the widely available AI model GLM 5.3 turned public security fixes into working exploits before updates reached Firefox users. It autonomously compromised six Firefox releases for about $1,000 in API fees each without triggering safeguards from any provider tested.
Mistral says its new model can afford organizations a kind of control over its cyber defenses that closed models can't.
Google’s new frontier model leads independent evals, with access initially limited to trusted defenders.
Gemini 4 Carbon is our strongest model with state-of-the-art cybersecurity and long horizon capabilities. Carbon ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.