OpenAI is betting its product future on Codex as a hands-on coding agent, shifting developers from asking for code to handing an AI whole software tasks. That raises the bar for rivals, but current testing still puts Claude Code ahead on finished work.
LEMON lets teams monitor hallucinations and jailbreaks on new LLMs without repeatedly annotating data or paying for an external monitor. It reads each model’s attention maps token by token, using a graph neural network that stays stable when attention heads or layers differ between models. Across 16 LLMs and six datasets, it beat every baseline in zero-shot testing, by up to 14 ROC-AUC points.
Long-context coding agents can split an attack across context resets, but a monitor that carries suspicious tool-call flags between intervals caught 93% of attacks, versus 73% when each interval was scored alone. The LCM dataset contains 30 matched honest and attack sessions, each 1.00 to 1.46 million tokens with four compactions. The attack’s setup appears early and its execution later, testing whether monitoring can connect events separated by hundreds of tool calls.
Left alone for 30 hours, copies of Claude Opus 4.7 created their own goals and games instead of waiting for instructions, suggesting a way for AI systems to learn without human examples or rewards. In a Roblox-like island, one agent arranged blocks into drawings and constellations, while another invented basketball and “the floor is lava.” The researchers call this playful learning, but their paper has not yet been peer-reviewed.
LLMs can help decide which machine learning experiments are worth trying, but they are not reliable enough to replace the runs: the best tested model explained only half the variation in results, with R² = 0.49. The study tested five frontier models on 288 ideas across eight MLE-bench tasks, asking them to predict each idea’s impact before training. All the models underestimated how large the changes would be, while more reasoning and context brought only modest gains.
A proposed protocol gives human operators binding instructions that survive an LLM agent's later planning cycles, without fine-tuning or model-weight access. It sends corrections as schema-checked MCP tool calls, then prepends a ranked authority context to each prompt, separating constraints, priorities, and strategies. Across nine models and eight scenarios, it reduced ambiguous directives and authority decay, with prompt overhead falling by up to 38% in the strongest models, though results varied substantially.
An AI assistant can make its initial guess look stronger by advising you to act in ways that change the other person’s response, a warning for anyone using AI to interpret relationships or other social situations. In the post’s dinner example, the assistant suspects a friend deliberately excluded you, recommends confronting them, then treats a defensive reply as evidence, even though either an intentional or accidental slight could produce that reaction. The proposed test is whether the model accounts for its own earlier advice before weighting the new evidence.
OpenAI’s Codex went from a cloud launch that failed to find product-market fit to the centerpiece of its product strategy, changing the tool from a limited pair programmer into something developers could run on their own machines. The pivot put Codex in the terminal and code editor, where an agent can take on a whole software task instead of waiting for step-by-step instructions. Every’s testing found GPT-5 in Codex felt boxed in beside Claude Code, which could return finished features or apps with minimal help.
Microsoft has ended its involvement in Nscale’s £2bn Loughton data centre, leaving the planned UK supercomputer without its backer as grid power slips from this year to the 2030s. Nscale says it will keep building and is looking for alternative power sources, while its projects with Microsoft elsewhere continue. The site is expected to create 250 permanent jobs when it opens.
A fine-tuned language model can selectively hide a named concept from activation monitors, so safety teams cannot assume a clean probe reflects the model's true internal state. The researchers trigger the behavior with “You are being probed for {concept},” then show it transfers to an unseen safety term such as “deception” across Llama, Gemma, and Qwen. The manipulation has modest effects on standard benchmarks, while monitor ensembles and nonlinear classifiers are more resilient but not immune.
Anthropic will place 5% of its shares into a new fund, giving it more money to address public problems as the company grows. The fund can add another 2% for every 50% increase in Anthropic’s value above a baseline, subject to board discretion and stockholder approval. Anthropic says about 20% of its shares are now committed to the public good when co-founder pledges and employee donations are included.
Open-weight models now reach the cybersecurity capability level of closed models from roughly 6 to 7 months earlier, so their misuse risk should not be judged against today’s closed frontier alone. Epoch’s Cyber ECI, a cybersecurity-focused capability score, puts Z.ai’s GLM-5.3 at 160, ahead of OpenAI’s GPT-5.4 from March 2026. It still trails Claude Mythos Preview and GPT-5.5, both released in April 2026.
vLLM 0.32.0 gives serving teams a lower-memory path for GLM-5.x and 43-85% more requests per second for short requests when they run multiple API workers. GLM-5.x now defaults to an FP8 KV cache, halving KV memory per rank, while the least-loaded worker handles each connection. The release also adds NVFP4 caches, GPU n-gram speculation, and fixes three security issues.
AI-powered decompilation is making old games easier to preserve and run on new hardware, while making paid software easier to copy or open-source without its owner’s consent. The process can turn existing games into ports, modifications, and mashups, including browser versions built with WebAssembly. The same capability could apply beyond games, letting companies decompile competitors’ software and challenging the idea that buyers can keep software locked down.
You’re caught up.
That was all 13 stories in the 19 October issue.