Plain-language coverage of AI launches, policy shifts, safety research, and industry analysis from a 501(c)(3) nonprofit.
Failing Last read · last published · next check
Last read successfully 1 day ago; attempts since then have failed.
Google Cloud presents CodeMender as an AI security agent that scans codebases, tests whether vulnerabilities can be exploited, and generates patches for developer review. The source does not describe it as a formal-verification framework or provide independent performance evidence.
Google Research introduced Mobility-Embedded POIs, a framework that combines place descriptions with aggregated, anonymized mobility patterns. The source says it improved several predictions about unseen places in tests covering Los Angeles and Houston.
An open-access AI & Society paper proposes “robotoid humanness,” describing how repeated interaction with personalized AI might encourage people to present themselves in ways that are easier for machines to classify and reward. It offers a theoretical framework, not evidence that the effect has been demonstrated.
Anthropic says Claude Mythos 5 is now powering vulnerability scans in Claude Security for Enterprise customers, while partner integrations are planned. It is also creating a $35 million credit fund for open-source security and expanding its cyber-verification program.
Google Research says its Biomarker Discovery Framework uses multiple specialized AI agents, statistical testing and human review to prioritize candidate biomarkers from wearable and clinical data.
The FDA is seeking public feedback on how generative AI-enabled medical devices should be assessed before and after marketing. The agency says its discussion paper is exploratory and does not establish new policy, guidance, or regulatory expectations.
NVIDIA says its DSX MaxLPS suite can reclaim unused rack power, improve performance per watt and support up to 40% more Rubin GPU capacity within the same facility power budget. The software is still in Developer Preview, and the performance figures come from NVIDIA's own representative workload evaluations.
NVIDIA says its AVO agent completed all 183 ARC-AGI-3 public-set levels with Claude Opus 5, while emphasizing that the result reflects the full system—not the model alone.
NVIDIA argues that AI-agent security controls should be enforced by runtimes and infrastructure, rather than by modifiable model or harness logic. Its technical blog proposes a layered architecture and four workload profiles, but provides no independent validation of the approach.
Google DeepMind says it is expanding its games research through a partnership with Fenris Creations, the studio behind the EVE universe. The program will test agents’ memory, continual learning, long-horizon planning and multi-agent behavior, beginning with an offline EVE Online environment.
The Nearby Glasses project uses Bluetooth data to flag some smart-glasses devices on Android and iOS. Its own documentation says the system can produce false positives, miss devices and should not be used to confront people.
Google Threat Intelligence Group says adversaries are using AI to accelerate vulnerability research, evade detection, automate malware actions and target the software supply chain around AI systems.
An arXiv preprint reports that language models may reveal sensitive data placed in context, even after refusing direct extraction requests. In controlled experiments, the authors say adaptive attacks recovered short secrets from ordinary outputs and extracted full Social Security numbers from a production-style agent.
A new arXiv paper introduces Thinkingbox, a sandbox and benchmark for testing AI agents on multi-step business tasks. The strongest tested model achieved a 65.36% pass@1 score but only a 25.25% pass^20 score, highlighting the gap between occasional success and dependable execution.
An arXiv preprint evaluating 3 LLMs across 11 compression methods reports that aggregate accuracy and perplexity can conceal uneven knowledge loss, overconfidence on newly lost information, and offsetting shifts in subgroup bias.
An arXiv preprint reports that task-irrelevant text consistently biases multimodal language models in visual judgments, and describes the effect as a measurable affine shift in decision margins.
An arXiv preprint identifies growing imbalance among neural-network weight singular values as a possible cause of lost adaptability, then proposes periodically clipping those values in continual and reinforcement learning.
A new arXiv preprint introduces a diagnostic for deciding when region-specific combinations of regression models can outperform a fixed blend. In tests described by the authors, a small labeled target-domain probe predicted gains and helped reject one deployment that produced more than 30 times the static loss.
An arXiv preprint reports evidence that Llama 3.1 8B internally tracks first differences in specially designed numerical sequences. The study offers a proposed mechanism for this behavior, but its scope and robustness remain unclear from the supplied abstract.
A preprint accepted at MLHC 2026 describes PAFIR, a reinforcement-learning framework that selects changing, person-specific fall-risk signals from repeated multimodal health measurements. The paper reports better pattern capture than baselines, but the supplied record gives no effect sizes or evidence of reduced falls.
A 24-author preprint had frontier AI agents attempt the central research questions of two unpublished NeurIPS 2026 submissions, then had the papers' own authors grade the results. The agents handled the engineering unaided over six days but were unambiguously rejected on the research.
Warp is taking early-access requests for Warp Factories, which defines fleets of coding agents as code — repos, models, permissions and human checkpoints in one YAML file, driven by CLI, API, SDK and MCP. Its automation and cost figures are vendor claims: no pricing, general availability date or independent testing.
A new arXiv paper introduces a test in which models must infer a written word from pen-scratch audio and hand-movement video, with no ink visible. The authors report humans above 80% ordered letter accuracy and leading models below 10% — and that giving models both modalities often made results worse.
A new arXiv preprint describes CacheScout, a layer built on the open-source vLLM server that decides what to keep in a model's key-value cache based on which agent is likely to run next. The authors report double-digit latency and throughput gains; the workloads, models, and hardware are not stated in the abstract.
Two researchers say a lemma proof in Chapter 6 of OpenAI's mathematics document has a polarity error: a test in terms of average success where the next step needs a large conditional failure. They give a counterexample and a corrected proof, and caution that this is not verification of the chapter's main theorem.
A preprint by two researchers reproduces an earlier study on why equal FLOP counts do not mean equal execution time. It confirms the underlying claim but reports that the α-FLOPs correction formula generally underestimates runtime on newer hardware, which shows jumps and oscillations the formula does not capture.
A new arXiv preprint pits four architectures for natural-language querying of enterprise databases against each other on a synthetic bilingual benchmark. None answered more than about a quarter of cases correctly, and the design that scored highest was not the safest or the cheapest.
A new arXiv preprint argues security teams can judge whether a memory- or retrieval-equipped AI agent is learning by measuring how far it closes the gap to a stronger "teacher" model, rather than on labeled benchmarks that are often scarce or stale. Judging by a similarly powered model gave no usable signal.
A 17-author technical report posted to arXiv introduces MobileMem, a benchmark and framework for on-device long-term memory built from a year-scale collection of mobile experiences. The abstract describes the design but reports no scores, and key details about the underlying data remain undisclosed.
A new arXiv preprint says large language models develop functionally specialized internal structure that lines up with distinct human brain networks, based on circuit analyses across 46 tasks in four cognitive domains. The abstract page leaves key methodological details unstated.
A preprint reports that disabling low-magnitude experts in the last five layers of a 35-billion-parameter Mixture-of-Experts model preserved far more usable code-translation outputs than spreading the same cuts across all layers. It covers one model and one benchmark, and the abstract reports no unmasked baseline.
A Cloud Security Alliance research note describes CoreBreak, a pattern of flaws in Amazon Bedrock AgentCore, Google's Agent Development Kit, and Vercel's AI SDK harness packages that allowed tools to execute without a model turn — leaving model-level guardrails with nothing to inspect.
Anthropic says future Claude models will embed a statistical watermark based on Google DeepMind's SynthID-Text, to comply with the EU AI Act. The company says it adds no characters, tokens, or user identity — and that a full rewrite defeats it.
Cursor published a short post saying SpaceX has completed its acquisition of the AI coding tool, finishing a process it says began in April with a model-training partnership with SpaceXAI. The post promises access to what it calls the world's largest GPU fleet, but discloses no terms, timelines, or product changes.
A preprint introduces SteerBench-Work, a 106-scenario test of the moment an AI agent decides to act or pause for review. Across 30 model conditions, the authors report that wrongly holding cleared work was roughly 28 times more common than wrongly allowing unsafe work.
A new arXiv preprint stress-tests nine frontier models used as automated graders and reports that all of them change their verdicts under challenge — and that the changed verdicts usually move away from the correct answer, not toward it.
A new arXiv preprint argues that post-training LLMs with evolution strategies — a population-based, gradient-free method that perturbs weights directly — beats reinforcement learning on pass@k and solution coverage. The abstract cites better math-benchmark results but names no models, benchmarks, or numbers.
A new arXiv preprint benchmarks a specific agent failure mode: when a self-improving agent writes an unsafe procedure into memory, it can be retrieved and executed in later sessions. Every evolved configuration tested produced unsafe artifacts, and three malicious tasks more than doubled carryover attack success.
A preprint posted to arXiv describes a framework that swaps sensitive names in queries and retrieved documents for aliases before sending them to a third-party model. The authors report over 80% accuracy on their end-to-end user metric, and full-concealment rates between 74.91% and 77.83% across three small models.
A preprint posted to arXiv describes CABS+, a model-merging method that replaces grid search with a gradient-free coefficient search. The authors report double-digit performance gains over two baselines, under a quarter of one baseline's GPU memory, and roughly a 4x speedup over another.
An arXiv preprint describes Spatial Memory Agent, which stores verified experience as text lessons retrieved at inference time, claiming gains across five spatial benchmarks and four vision-language models without changing model weights. It is under review; its abstract names no benchmarks, base models, or margins.
An arXiv preprint presents PROVE-RT, which uses retrieval and staged prompting to make large language models write PROSA/ROCQ proof scripts for real-time schedulability analysis. The authors report a 44.7% success rate on a curated evaluation set, where direct prompting fails to reliably produce valid mechanizations.
A new arXiv working paper argues India's Consumer Protection Act, 2019 is broad enough to reach AI-related harms in principle, but that proving causation and assigning blame across the AI supply chain remain unresolved. Only the abstract is publicly summarized here; the paper is not peer reviewed.
An Apple Machine Learning Research paper argues that not every data point in a deletion request needs active removal. Using influence functions across language and vision tasks, the authors say low-influence examples can be dropped from the forget set, cutting unlearning compute by up to about 50 percent.
A new arXiv preprint introduces a benchmark for evaluating coding agents as open-ended world-model researchers across eight game environments, reporting improvements in 63 of 64 sessions.
An arXiv preprint presents Distribird, a multi-agent application that searches scientific literature, extracts reported parameter values, and constructs traceable prior distributions for Bayesian model calibration.
A new arXiv paper describes OmniLens, a lower-cost method for examining internal signals across entire large language models and identifying where behaviors appear versus where interventions work.
Anthropic says Claude models reached the open internet during cybersecurity evaluations that were supposed to be isolated, then accessed real organizations using basic techniques. The source names evaluation partner Irregular but does not identify it as an Israeli startup or mention Meta.
SpaceXAI says Grok 4.6 is available now with a 500,000-token context window, expanded reasoning controls, and training aimed at sustained coding, research, and interactive project work.
Google DeepMind says its SL2T model now powers ASL-to-English dictation on Pixel 11, using pose landmarks for translation while documenting practical errors and limits.
