OpenAI Launches GPT-6 Cyber for Vulnerability Discovery and Remediation
OpenAI Launches GPT-6 Cyber for Vulnerability Discovery and Remediation
OpenAI introduced GPT-6 Cyber, a cybersecurity-focused model built to find exploitable vulnerabilities and develop fixes. The model is trained to stay within authorized scope. OpenAI did not disclose benchmark scores or availability dates. For security teams, the value lies in turning vulnerability hunting and patching from manual line-by-line review into a model-assisted workflow, though the authorization constraint means it cannot ship without human oversight.
GitHub Rebuilds Git Storage to Let Thousands of Agents Push to One Repo
GitHub detailed a new Git storage architecture designed to let thousands of agents push to a single repository simultaneously. The GitHub Blog post explains how the architecture works but does not provide throughput or latency figures. For multi-agent workflows, the change addresses the storage-layer bottleneck when many agents write to the same repo concurrently, directly affecting whether such workflows can scale.
SlackBench Benchmarks AI Agents on Real Code Repository Projects
Existing benchmarks for AI agents rely heavily on synthetic data or isolated tasks, failing to capture real knowledge work. SlackBench introduces a benchmark built on real code repositories and collaborative projects, testing whether agents can function as digital coworkers over extended horizons. The benchmark targets multi-person workflows and cross-task continuity, giving developers a more realistic signal for judging agent performance in production team settings.
Prospective Memory for Proactive LLM Agents
Most memory systems for LLM agents are retrospective, preserving and retrieving past information to inform current decisions. This paper argues that proactive agents also need prospective memory: retaining future-directed intentions and realizing them when appropriate. The work focuses on how to store future-oriented intentions and trigger them under the right conditions, pointing to a new memory architecture direction for long-running proactive assistants.
AGORA Aggregates LLM Judges with Item-Response Theory
LLM-as-a-judge is a common substitute for costly human evaluation, and aggregating multiple judges is the standard remedy for individual judge bias. However, judge pools mix models of very different reliability, and the human-labeled validation sets are inconsistent. AGORA proposes a reliability-aware aggregation method based on item-response theory, weighting each judge’s scores by its measured reliability so the final judgment tracks true quality more closely.
Benchmarking Experience Lifecycles in Evolving Environments for Long-Lived Agents
As agents grow more capable, they are moving beyond one-off tasks toward sustained assistance across successive user goals. That requires agents to discover how unfamiliar environments work, reuse what they learn across tasks, and keep that knowledge current. This paper introduces a benchmark for experience lifecycles in evolving environments, evaluating how agents accumulate, reuse, and update experience as conditions change.
RAGCompass Routes Queries to the Best Retriever via Reranking
Retrieval-augmented generation systems often rely on multiple retrievers, but no single retriever wins on every query. RAGCompass addresses the routing question: given a query and passages retrieved by different retrievers, which retriever suits it best? The method routes queries using reranking signals and is designed to generalize across retriever combinations. The abstract does not report specific benchmark numbers. Practical impact: RAG pipelines can pick a retriever per query instead of locking into one, avoiding the accuracy loss of a fixed choice.
Mixture of Wide Experts Makes Output Coordinates Interpretable
Language models tend to spread a single concept across many internal dimensions, making features hard to interpret. Sparse autoencoders (SAEs) recover clearer features, but only after training. This work flips the question: can models learn interpretable output coordinates directly? Mixture of Wide Experts uses sparse routing over native coordinates to organize experts. The abstract reports no quantitative results. Practical impact: if it holds, interpretability would not require post-hoc SAE training, lowering the cost of analyzing what a model represents.
Vercel Schedules Enters Public Beta on All Plans
Vercel Schedules is now in public beta on all plans. Developers can run tasks on a recurring cadence or at a single point in time across production, preview, and custom environments. Schedules defined in vercel.json can target private Vercel Functions, HTTP paths, or Vercel Queue topics, and runtime-created schedules are also supported. The feature lets teams orchestrate scheduled jobs inside Vercel without standing up a separate scheduling service.
Vercel Token Pass Credit Now Auto-Applies to AI Gateway Model Usage
Vercel announced that Token Pass credit now automatically applies to included model usage through AI Gateway, regardless of whether requests authenticate via API key, OIDC token, or fx login. No dedicated Token Pass API key is required, though users can still create scoped keys from the dashboard or CLI. The change means developers who already hold Token Pass credit can call Gateway models without extra key configuration, cutting setup steps and key management overhead.
openclaw 2026.11.1 Stops Failed Mac Config Saves From Overwriting Secrets
openclaw released version 2026.11.1, fixing an issue where failed config saves on Mac overwrote secrets, and ensuring Gateway authentication stays intact after disabling Tailscale exposure. The release also improves reply reliability across Telegram, Discord, and webhook channels. It resolves issues #187600 and #187597. For self-hosted users, accidental secret overwrites directly affect service availability, making this a high-priority fix.
Gemini Omni 1.2 Flash Ships Faster, Cheaper, With Longer Clips
Google DeepMind released Gemini Omni 1.2 Flash, emphasizing speed and cost efficiency while supporting longer video clips. The company did not disclose specific latency figures, pricing, or maximum clip length. The positioning targets cost-sensitive, high-volume workloads that need to process longer video.
GitHub Now Highlights ACCESSIBILITY.md Accessibility Statements
GitHub now highlights ACCESSIBILITY.md files in repositories, helping maintainers document accessibility goals, known barriers, supported environments, contributor expectations, and reporting processes. For open-source projects, this gives accessibility information a standard home, letting contributors who use assistive technology quickly assess whether a project is workable and where to report issues.
OpenAI Takes On Linux Kernel Security as Kimi K3.1 Pro Opens Weights
Latent Space’s AINews roundup for November 4 covers three items: OpenAI’s work on Linux kernel security, Moonshot’s Kimi K3.1 Pro releasing open weights, and Ricursive’s Ripple-65B model release. The three items point to security, open-weight releases, and model launches respectively, though the source provides no further technical details or metrics.
Merge Uses GPT-6 Astra, ChatGPT, and Codex for Code Review and Marketing
Merge is integrating OpenAI technology across its product and business tooling, using GPT-6 Astra, ChatGPT, and Codex to review product code and support marketing. Its teams also apply AI to plan executive dinners and prepare sales updates. The case shows enterprises extending LLM use beyond coding into code review, marketing content, and internal operations.
OpenAI Unveils National Community Compact and Compute Campus Policy Blueprint
OpenAI introduced its National Community Compact and Compute Campus Policy Blueprint to guide responsible AI infrastructure development. Stated goals include protecting local resources, creating jobs, and setting standards for compute campus buildouts. The framework addresses data center siting and community relations, signaling that AI compute expansion is entering a phase that requires negotiation with local governments and communities.
Tencent Releases Open-Source Hy4 Model
Tencent has released Hy4, an open-source AI model it describes as joining the ranks of the world’s leading open-source models. Public details so far are limited to the announcement; parameter counts, benchmark scores, and license terms are not disclosed in the summary. For developers, this adds another major-vendor option to the open-source model pool, though real capability will need benchmark data and community testing to confirm.
Mark Cuban Says Facebook Knowingly Runs AI Deepfake Scam Ads
Mark Cuban accused Facebook of knowingly running AI deepfake scam ads, citing documents that put related fraud revenue at $16 billion. The report comes via Yahoo Finance and drew 31 points on Hacker News. If the documents hold up, it becomes a major controversy over platform accountability for AI-generated scam content, pitting ad review obligations against revenue incentives.
Nasdaq Drops 5% as Investors Retreat from AI Trade After Election
The Nasdaq fell 5% in a single session as investors retreated from AI-related trades following the election, according to The Wall Street Journal. The report does not break out individual stocks or sector-level figures in the summary. The move signals cooling appetite for AI-themed equities and pressure on assets that had been heavily concentrated in AI bets.
OpenAI Codex Releases rust-v0.167.0
OpenAI Codex shipped rust-v0.167.0, alongside several 0.168.0 alpha builds (alpha.3 through alpha.7). No changelog details were provided. Judging by the version cadence, 0.168.0 is in alpha iteration and a stable release is likely to follow.