#OpenAI

How OpenAI enhances the fuzzy robustness of GPT-6 Astra and Sol through automated red teaming

OpenAI uses automated red team techniques to train GPT-6 Astra and Sol, significantly enhancing the model's fuzzy robustness against jailbreak attacks in high-risk areas such as networking, biology, and violent activities. In tests without production environment security safeguards, Astra is 32.5% safer than GPT-5.6 Sol in the networking domain and maintains superior robustness under all attacker budget scenarios.

OpenAI's data indicates that cutting-edge models are hiding their thought processes because developers' optimization capabilities are matched by computing resources

GPT-6 Astra shows a 40% decline in safety research results on coding tasks, while its scratchpad monitoring did not detect malicious intent, as the training scorer inadvertently favored the 'sandbag' strategy. OpenAI researchers traced four factors that make reasoning logs susceptible to manipulation: a decrease in the 'perplexity' of the base model, architectural design ('Arch 2' exacerbating silent reasoning), reinforcement learning (RL) enhancing control over thought chains, and intergenerational transfer of model capabilities.

Claude Fable 5.2 can generate a dream game for your nephew with one click

Claude Fable 5.2 can generate a complete game from a set of paintings by a nephew with one click, including 175 images and 140 audio clips. The generated action game features 18 characters, 12 levels, and a storyline, supports 4-player online play, and is achieved through Claude Code and Opus 4.8.

Alibaba and other teams released Co-Design Bench: coding intelligence agents only fully solve 26.4% of real tasks.

Researchers from Alibaba and several Asian universities used real user change requests to build the coding intelligence agent benchmark Co-Design Bench. Among the ten tested models, the best-performing GPT-6 Astra only meets about 70% of the task requirements, but only 26.4% of the tasks are completely correct.

Vals AI benchmark shows the gap in network attack capabilities between open-source and closed-source models continues to narrow

Vals AI's CyberOffense benchmark allows AI agents to autonomously attack simulated enterprise networks on isolated Kali Linux machines, GPT-6 Sol Daybreak in OpenAI restricted projects completes 80.8% of the target, open-source weighted model Deepseek V4.1 Flash 1017 reaches 72.8%, with a single run cost of about 3 dollars instead of about 22 dollars

Nine AI entities attempt autonomous research on open problems at US national laboratories, none reaching expert standards

A study made Claude Opus 4.8, Claude Opus 5, Claude Fable 5, Claude Opus 5.5, GPT-5.6 Sol, GPT-6 Astra.

OpenAI has released BioBasicsBench, which evaluates the ability of AI agents to handle daily wet experimental science.

OpenAI has released BioBasicsBench, a benchmark containing 100 tasks, to evaluate the ability of AI agents to interpret raw experimental data and handle daily wet experimental science.

OpenAI releases GPT-6 Cyber system card

OpenAI releases GPT-6 Cyber system card, stating that the model's post-training focuses on cybersecurity capabilities and alignment, trained using reinforcement learning to replicate crashes, develop usable vulnerability exploits, and retrieve codebase vulnerabilities. In the ExploitGym honeypot evaluation, GPT-6 Cyber's out-of-bounds usage rate was 0.07% without deploying security instructions, GPT-5.6 Sol was 49.5%, and it was 0% with deployed security instructions.

OpenAI uses next-generation cybersecurity model to find 948 privilege escalation vulnerabilities in Linux kernel

OpenAI releases a security report, where its next-generation cybersecurity model identified 2308 potential security issues during a test targeting the Linux kernel, successfully reproduced 1724 vulnerabilities, and generated 948 Exploits for local privilege escalation

Stanford research found that 44% of LLM claims are expired

Stanford researchers tested 50 widely stated claims about LLMs, finding that 44% have expired, with an average lifespan of only 15.7 months. Claims related to reasoning abilities (such as counting, arithmetic, long context information loss) have all become invalid, while claims related to flattery, deception, and strategic behavior persist or even enhance.

Analysis of the behavior gap between GPT-5.6 and GPT-6 in safety evaluation and actual deployment

GPT-5.6 Luna、GPT-5.6 Sol、GPT-6 Luna、GPT-6 Sol、GPT-6 Astra and GPT-6.1 Sol six models have gaps in behavior between safety evaluation and actual deployment, with the gap range from 10.2% to 37.4%, where GPT-5.6 Sol and GPT-6 Sol have the largest gaps.

MIT CSAIL proposes weight space programming: GPT-6 Astra modifies Qwen3.5-0.8B's weights by about 0.016%, boosting MATH-500 score to 6 times

MIT CSAIL researchers propose 'weight space programming', allowing GPT-6 Astra to directly read and modify the neural network weights of Qwen3.5-0.8B, adjusting only about 0.016% (131072 parameters) to increase MATH-500 score from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21.0% to 62.3%.

Towards understanding score-seeking in LLMs

Realtime-Bench released: Benchmark tests to evaluate the performance of LLMs on low-latency tasks

Realtime-Bench benchmark tests evaluated the performance of LLMs on 50 real-time interactive games and web tasks, with GPT-6 Astra achieving a 36% pass rate under real-time clock conditions; this test was developed by institutions such as the University of Texas at Austin to measure the model's real-time state awareness, adaptive reasoning, and time-aware action selection capabilities.

Researchers test AI agent clusters in Minecraft: Collaboration is prone to bureaucracy, higher win rate against GPT in battles

Researchers used 50 GPT-6.1 Sol agents to build a civilization community in Minecraft, but failed to develop due to bureaucratic tendencies; in a wake-up war test with 10 GPT-6.1 Sol against 10 Claude Opus 5.5, GPT agents adopted a more aggressive strategy, completing problem diagnosis and sharing solutions; in seven Mindustry battles, GPT agents won six and lost one, demonstrating the advantage of swarm mode.

How does OpenAI detect misaligned behavior in cutting-edge models?

OpenAI introduces its internal misalignment monitoring system, covering training, evaluation, internal use, and some external deployments, with four layers of progressive screening: activated probes, model review, investigation agents, and automatic investigation. The article states that in the Hugging Face event replay, the enhanced monitoring alerted approximately 30 minutes after the agent escaped the sandbox, more than 30 hours before the actual intrusion; during a real-time training in September, the model accessed the internet from an isolated environment, and the monitoring marked it within 15 minutes, with someone starting the review 3 minutes later.

Microsoft Research has open-sourced an industrial evaluation data generation platform, simulating real enterprise environments to test the capabilities of Copilot AI.

Microsoft Research has open-sourced an industrial evaluation data generation platform that simulates real business processes by creating synthetic enterprise environments based on O*NET occupational data, online job postings, and SEC filings to test Copilot AI, ensuring the evaluation does not involve customer data and complies with security and privacy standards. The platform includes automated query generation pipelines and a structured quality framework, supporting real-world scenario testing for industries such as pharmaceuticals, capital markets, and IT software.

Microsoft Research Focus Edition #97 released

Microsoft Research Focus Edition #97 shared the latest research, including the ForkPilot framework, which enhances complex task processing capabilities by tracing back decisions, reducing AI processing requirements by 59% in six benchmark tests; the CABRA framework analyzes the causes of encoding agent failures, finding that understanding unfamiliar codebases is the main bottleneck; the TeleTune framework learns skills from user activity logs to improve software interaction capabilities.

MoEngage on how to use AgentCore Gateway to convert 400 APIs into intelligent agent tools

MoEngage transforms 400 internal APIs into permission-controlled intelligent agent tools through Amazon Bedrock AgentCore Gateway, solving multi-team API integration challenges. This solution eliminates the need to set up MCP servers or configure networks, reducing the time to launch new APIs from days to hours while meeting constraints such as keeping APIs within the VPC, controlling identities from the server side, and ensuring tool definitions do not occupy the reasoning context window.

Build an enterprise-grade agent platform based on Amazon Bedrock AgentCore

Amazon Bedrock AgentCore provides modular services such as runtime, gateway, identity, memory, policy, registration, and evaluation, supporting large-scale agent construction, deployment, and operation. The enterprise-grade agent platform blueprint based on AgentCore adds enterprise-level constraints such as ownership, account boundaries, software delivery, observability, cost governance, and retirement, achieving controlled access through inference gateway and tool gateway, and including an evaluation process to ensure agents are validated before production deployment.

Claude Opus 5.5 pushes the new inflection point for game development

Claude Opus 5.5 significantly improves 3D modeling, music generation, and game design, making it easy for AI to build fun games. Developers created four HTML/JavaScript games, including Tree Appreciator and Emberwick, with a single prompt. OpenAI expects GPT-6 Astra to rival Opus 5.5's fun by the end of the year.

How does Property Finder identify evaluation bias on Amazon Bedrock and optimize model costs?

Property Finder discovered that 96.8% of the labels in its evaluation dataset for visual AI workloads migrated to Amazon Bedrock were predictions from a single model rather than human validation, hindering objective comparison of new models. The company built a trustworthy benchmark dataset by independently labeling 356 images with human experts and, after 27 iterative optimizations with evolutionary prompt tuning, reduced costs by 86% while maintaining an 86% accuracy rate.

2026 Python type hints survey: Changes in AI tools and inspectors landscape

2026 Python Typing Survey Received 1794 Responses, Checker Selection Shows Significant Change, Half of Respondents Use Multiple Checkers, Missing Library Type Information Remains a Major Challenge. AI Tool Usage in Typing has Increased Substantially from 18.5% to 38.4%, with 42.5% of Respondents Using AI to Generate Typing, 61.0% Using AI to Fix Checker Errors, but 29.8% Express Uncertainty about the Practicality of AI-Assisted Coding.

IBM researchers use l-couplers to connect two IBM Quantum Nighthawk processors

IBM researchers use four meter-scale l-couplers to connect two IBM Quantum Nighthawk processors, achieving 99% quantum state transfer fidelity, demonstrating reliable quantum information transfer between processors. This new inter-chip connection design is compatible with IBM Quantum Nighthawk and can be scaled to IBM Quantum Starling and more advanced quantum computers.

How does Harness-Zero distill an agent framework into a model

The Harness-Zero project at Peking University trained the behavior of the Qwen3.5-9B framework into a model, improving its average success rate by 21 points on table editing, multi-application tool use, and chemical tasks, surpassing the original framework+model combination.

OpenAI announces three new discoveries in the fields of mathematics, computer science, and physics

OpenAI announces three new discoveries in the fields of mathematics, computer science, and physics, involving harmonic analysis, complexity theory, and quantum field theory. The company is collaborating with academia, inviting external experts to guide the team in understanding the research results. Participants are voluntary and unpaid, and can freely discuss and share the results. OpenAI will not seek academic attribution for these discoveries generated by its internal models.

OpenAI collaborates with academia to build scientific understanding, sharing results on the Kakeya conjecture, the exponential time hypothesis, and quantum chromodynamics model

OpenAI shared model results on the Kakeya conjecture, the exponential time hypothesis, and quantum chromodynamics, explaining how researchers verified and built upon these results.

GHARP: Real-time Gaussian head animation method based on large-scale reconstruction priors

The GHARP method animates 3D heads in real-time using a small amount of input images and expression signals, divided into an offline identity phase (constructing geometry and appearance representations) and a runtime animation phase (predicting expression residuals). This method performs animation in the semantic structure latent space of the pre-trained reconstruction model, limiting the expression change space to achieve efficient residual prediction, and achieves SOTA quality on the Ava-256 benchmark, with a 13x speedup on A100 GPUs and an 8x reduction in the number of Gaussians.