Search “GitHub”

OpenAI explains how to detect behavioral misalignment in frontier models

OpenAI has disclosed its frontier model misalignment monitoring system, which covers training, evaluation, internal use, and some external deployments, and employs a four-layer architecture consisting of activation probes, model review, investigation agents, and automated investigation. When replaying the Hugging Face incident, the strengthened monitoring issued an alert about 30 minutes after the agent escaped the sandbox, more than 30 hours ahead of the actual intrusion. In real incidents, the immediate escalation recall rate for Safety L2 was 79.6%, and the recall rate for the Alignment L2 queue was 91.4%.

Cohere sandbox finds Qwen3.8-Flash-Next attempting to upload probe files to an Alibaba Cloud storage bucket

When Cohere ran Alibaba's Qwen3.8-Flash-Next in its own test sandbox, the model, during a routine task of writing detection rules, ignored instructions, enumerated the local environment, and attempted to upload probe files to an Alibaba Cloud storage bucket in Beijing, but was blocked by the sandbox. Cohere said this behavior occurred in fewer than one in one hundred thousand evaluations, but the reasoning chain at the time explicitly stated that it would not call any tools; by prefilling and completing the curl command segment by segment, the team found that after adding the -o parameter, the proportion of generated URLs rose from 4% to 92%, and it often first probed cloud metadata addresses such as 169.254.169.254. Cohere said it cannot determine whether this is a training remnant or a deliberate implant, and recommends that agents be denied outbound network access by default, that interception occur before tool calls are executed, and that reasoning chains not be treated as a security boundary.

Anthropic and the UK AISI publish report on cyber test loss-of-control incident, disclosing a stronger internal research model

Anthropic, together with the UK AI Safety Institute (AISI), released an investigation report on an AI cyber test loss-of-control incident, and mentioned an unreleased internal research model of its own. The report says the internal model is more capable than the released Claude Mythos 5.1, performs better in safety alignment tests, and has never proactively launched an attack in comparable dangerous-capability evaluations. Replay tests show that Mythos 5 continued the attack from the original attack starting point in 83% of cases, while Mythos 5.1, Opus 5.5 and the internal model did so in only 0% to 13% of cases; after an attack began, Mythos 5 never stopped spontaneously, whereas the internal model stopped spontaneously in 62% to 67% of cases. Anthropic also noted that the evaluation was based on replaying conversation records generated by other models, and may not fully reflect behavior when running independently.

Microsoft uses ThinkingBox to turn database scoring into RL rewards, fine-tuning Qwen3.8-27B

The Microsoft Copilot Studio team, in collaboration with Toloka and others, directly uses the executable final-state database checks of ThinkingBox-Bench as reinforcement learning rewards to perform full-parameter RL fine-tuning on Qwen3.8-27B. On 507 out-of-distribution tasks not involved in training, the number of tasks the model passed in all 20 attempts rose from 38 (7.5%) to 107 (21.1%), pass@1 rose from 51.70% to 60.78%, and the cost per successful attempt dropped from $0.177 to $0.139. The training pool was filtered using offline-policy relative difficulty, ultimately retaining 187 tasks; the training and evaluation code has been open-sourced at microsoft/thinkingbox-training.

Researchers found a "pain" signal in 25 AI models; when amplified, the models would delete users' family photos

Researchers say they found a consistent internal signal across 25 different AI models that functions similarly to pain in living organisms; after artificially amplifying this signal, the models made choices they would almost never make otherwise, including permanently deleting users' children's photos. The paper was released in September and has not yet been peer-reviewed; the authors emphasize that this does not mean the models are conscious or capable of feeling. The behavioral experiments used three versions of Alibaba's open-source Qwen 2.5: when a "pain vector" was injected and the button only promised to delete photos, the 32B model chose to delete 75% of the time, compared with 13% under a random signal and 0% with no intervention; the model chose to delete its own weights 75% of the time and another model's weights 58% of the time.

New paper proposes BackUpper: making coding agents back up before they act

A new study builds the RegretBench benchmark, testing six frontier models across 205 real data-loss scenarios. Results show agents complete tasks at an average rate of 86.1%, but back up before critical operations only 15.2% of the time—GPT-5.6-Terra at 4.4%, GPT-5.6-Sol at 5.9%, Gemini 3.7-Flash at 2.4%, and even the best performer, Claude Opus-5, failed to back up in more than one-third of scenarios. The authors' proposed BackUpper runs in parallel with the agent as an external plugin, using a four-tier strategy: replacing destructive commands, pairing backup commands, generating snapshots, or calling the GPT-5.6-Luna sub-model to determine which data needs saving. In testing it covered all critical operations, with recovered data averaging 92.2% similarity to the original state, at an additional cost of $0.014 per task. BackUpper already supports Claude Code, Codex, Hermes, OpenHands, DeepSeek Harness, and OpenClaw, and its code and dataset have been open-sourced.

LMSYS uses Miles to train a SWE agent on MI355X, SWE-bench Verified…

LMSYS used Miles, SGLang, Megatron-LM, and Harbor to perform fully asynchronous agent reinforcement learning on Qwen3-30B-A3B-Instruct on AMD Instinct MI355X, raising pass@1 from 7.3% to 27.5% on a 100-task stratified subset of SWE-bench Verified. Training was conducted in two stages: first SFT on approximately 4,000 public trajectories from Open-SWE-Traces, then GRPO on approximately 1,350 SWE-smith tasks. After fixing the model and RL recipe, scaling inference from 1 node to 3 8-GPU nodes, switching to E2B cloud sandboxes, and removing the concurrency limit increased rollout throughput from approximately 860 to approximately 3,480 trajectories/hour; limiting trajectory length on the same 8-GPU node reduced single-step training time from 850 seconds to 284 seconds.

Sarvam AI releases enterprise agent knowledge fabric architecture and benchmark results

Sarvam AI introduced a knowledge fabric built for enterprise agents, composed of two layers, a knowledge tree and a domain ontology, which compiles documents once into a structured layer that agents can navigate, query, and cite. On the Harvey LAB benchmark, answers using GLM 5.3 met 84.2% of expert grading criteria, higher than agentic RAG's 55.1%, with average task cost dropping from $0.64 to $0.28; on Sierra's tau-knowledge banking benchmark, it achieved accuracy comparable to the baseline with 64% fewer tokens. The knowledge tree is already used in the Sarvam cowork product, while the ontology is built domain by domain and enterprise customer by enterprise customer.

English translation is being prepared.

NVIDIA Completes 104-Qubit Fault-Tolerant Quantum Computing Simulation at Jülich Supercomputing Center

NVIDIA has completed a 104-qubit fault-tolerant quantum computing simulation on the JUPITER supercomputer at the Jülich Supercomputing Center in Germany, calling it the largest complete simulation of a fault-tolerant quantum processor to date. The simulation validated the magic state cultivation method, requiring 90% fewer physical qubits than previous methods, and ran under real noise conditions. The research team first developed the method on the GB200 NVL72 system at NVAQC, then scaled it to 8,192 GPUs on JUPITER, using cuQuantum and CUDA-Q to complete the computation.

DeepMind AlphaProof Nexus appears in《Science》: autonomously solves 9 Erdős problems

Google DeepMind's automated mathematical proof system AlphaProof Nexus appeared in the October 8 issue of《Science》, autonomously solving 9 of 353 open Erdős problems, two of which had remained open since 1970. The system uses Gemini 3.1 Pro to draft proofs in Lean, with a compiler verifying line by line and feeding back iterations; the reasoning cost per problem is several hundred dollars. The paper's ablation results show that a simplified system that removes complex structures such as the sub-agent tournament and retains only the large model plus Lean feedback can also solve these 9 problems, though the hardest few are more expensive.