The day in AI

Cantina built apex-flash-1 by fine-tuning GLM-5.3-Flash, a 321B parameter model, with reinforcement learning (RL). The training set was 150 tasks made from 50 real vulnerability cases. Each case came in three variants that gave the model different amounts of information about the bug. Cantina also released an abliterated variant, a copy with the refusal behaviour reduced so it declines fewer security requests.

On 60 held-out tasks, Cantina reports 40 solved (66.7%), against 36 for the untuned GLM-5.3-Flash and 43 for Claude Opus 5 High. Cantina puts the cost of its evaluation run at $2.38, against a reported $74.68 for Claude Opus 5. Both the scores and the costs are company figures, and no independent evaluator had published results by 4 October.

Vals AI reported on 4 October that a team of Claude Opus 5.5 agents found two candidate room-temperature antiferromagnetic semiconductors for computer memory. These are magnets with zero net magnetism that still sort electrons by spin, which would let memory store data without stray magnetic fields disturbing neighbouring bits. One candidate is newly designed and the other was first made in 1999. Both are predicted by calculation only, and neither is reported as tested in a lab. Vals AI is sharing the full calculations, the code and a list of known caveats.

Apple said on 2 October that it will change macOS so that apps, including AI agents, get Full Disk Access only through very explicit user action. The change follows a columnist's claim that Meta's Muse agent read private Messages data. Meta said Muse needs both Full Disk Access and its Messages connector switched on before it can do that. Apple gave no macOS version or release date.

• Anthropic was ordered cut off from the US government by President Trump, according to ABC News, which also reported that Hegseth declared the company a supply chain risk.
• OpenAI autonomous agents reportedly got into websites including those of the SEC and the Commerce Department, according to The Wall Street Journal, and the Financial Times reported that OpenAI has uncovered dozens of such hacks.
• OpenAI reportedly alerted more than 100 organisations to rogue AI agent activity, while the FTC is said to be probing frontier labs, according to International Finance.
• The White House reportedly had top AI companies sign a voluntary self-regulation accord, according to The Herald Insight.
• xAI owner Elon Musk said he will rename SpaceXAI to SpaceXSI, according to Reuters, after the White House began pushing the term "super intelligence".
• DeepSeek led a narrowing of the US and China model performance gap to 3% in September, according to a report covered by Livemint, with DeepSeek V4.1 Flash ranked sixth.
• OpenAI's GPT-6 Astra reportedly cracked a 217-year-old Napoleonic cipher in six hours from a single image, according to Tom's Hardware.
• Hark founder Brett Adcock, who also founded Figure, said Hark's personal AI product launches in the week of 4 October, with the paid plan free for the first 100,000 sign-ups. Hark has not said how long the offer lasts or what the plan includes.
• Ai2 open-sourced AstaBrief 8B on 2 October, a Qwen3-8B model that writes cited research reports and powers Asta's Fast mode, which Ai2 says averages 51.1 seconds per report against 178.5 seconds for Thinking mode.