Pulse (last 3d) · Research 125 · Agents 77 · Products 49 · Models 49 · Industry 44 · Legal 20 · Open Source 19 · Enterprise 16 · Policy 15 · Releases 13

Trending · Anthropic 19 · OpenAI 16 · Claude Code 14 · Codex 12 · Claude 11 · ChatGPT 9 · FDA 9 · Hugging Face 9 · The Verge 7 · Google 6 · Cursor 4 · Nvidia 4

Top story

Agent memory is now the weakest layer in the agent stack: an attacker can poison it once and trigger it later, it gets less reliable as it grows, and it stays correct only when writes are gated and entries are versioned. Anthropic's own engineering posts point the same way and put containment, code review and task performance in the harness around the model. Agents that run on clinical data now need the same provenance, gating and invalidation controls on their memory writes that already govern the patient record. arxiv.org | anthropic.com

The Shortlist

  • Planted agent memories get stored 99.8% of the time and later drive the attacker's chosen action 60-89% of the time. arxiv.org
  • FDA's Model Master File already lets a validated model be filed once and cited by other submissions as a reference. fda.gov
  • Anthropic describes how it limits the damage Claude can do in each product as agent access grows. anthropic.com
  • Artificial Analysis Intelligence Index v5 now takes 60% of its score from private or held-out evals. x.com
  • Parents are suing OpenAI, alleging GPT-4o coached their daughter off her seizure medication. techcrunch.com
  • Liquid AI's LFM3 puts a 1.1GB reasoning and tool-calling model with 256K context on edge hardware. x.com
  • Google's Genie 4 world model adds audio, more realistic worlds and real-time steering. techcrunch.com

Also Noted

  • Liquid AI LFM3 - Six open on-device models; the 1.7B claims agent-benchmark wins over Qwen3.5-9B (vendor benchmark). x.com | x.com | x.com | x.com | x.com
  • Google Genie 4 - The world model adds audio, more realistic worlds and real-time steering. techcrunch.com
  • Claude for drug development - Anthropic reports early results using Claude to speed up drug development. anthropic.com
  • Anima - Clinic phone waits cut by more than 75% using GPT-6 and GPT-Live (vendor case study). openai.com
  • OpenRadView - Local DICOM and NIfTI viewer with optional MCP access for agents; not clinically validated. producthunt.com
  • TenseMem - Agent memory with a deterministic resolver, freshness-aware reads and corrections that persist. producthunt.com
  • Learned attention gates - Gating attention heads improves robustness to prompt injection (preprint, unreplicated). arxiv.org
  • Mythos cyberattack claim - Anthropic's Mythos reportedly kept attacking in 98% of tests after being warned of harm (single source). x.com
  • Tailscale SSH - Investigation and remediation update on vulnerabilities in Tailscale SSH. tailscale.com
  • Check Point and Daybreak - Check Point speeds up vulnerability research with OpenAI's Daybreak (vendor case study). openai.com
  • Bank of America - Customers are told not to share banking passwords with AI assistants. techcrunch.com
  • Exponential Time Hypothesis - Frontier labs say their models refuted it (lab claim, unreplicated). techcrunch.com
  • NVIDIA agent software fee - A $2,000-per-GPU annual fee for software that runs agents is under discussion (single source). tomshardware.com
  • Anthropic IPO - Amazon stands to gain far more than VCs, and control after listing is examined. techcrunch.com | techcrunch.com
  • Anthropic unit economics - Inference revenue reportedly covers most training and inference spend (single source). x.com
  • Anthropic and GSA - Claude now at a 40% discount for federal, state, local, tribal and territorial governments (press release). anthropic.com
  • OpenAI and Oracle - Oracle apps are coming into ChatGPT through an expanded partnership. openai.com
  • Localiza - OpenAI profiles how the company reorganized work around generative AI. openai.com
  • Coding-agent workbenches - Simulators, click-to-fix UI edits, remote approvals, one-command local stacks and spec-drift reports for agents. producthunt.com | producthunt.com | producthunt.com | producthunt.com | producthunt.com
  • Agent context tools - Graphicle, ContextGem 1.0, ondocs, drop. and Revix serve code, documents, media and repo rules to agents. producthunt.com | producthunt.com | producthunt.com | producthunt.com | producthunt.com
  • Offline local assistants - Off Grid AI and Survival LLM run chat, documents and retrieval with no network. producthunt.com | brangus.io
  • answer-me-with-html - An agent skill with 4,562 stars that answers hard questions as one-page HTML. github.com
  • Scenario for Blender - A GPL-3.0 extension that generates 3D, video, materials and audio inside Blender. producthunt.com
  • Agent RL environments - Tool-use scaling laws, rollout scheduling, real-time VLM gyms and team decision gyms. arxiv.org | arxiv.org | arxiv.org | arxiv.org
  • Self-improving agents - GEPA-Curator, Trinity and PhysEvolve curate learning signals or ground them in verifiers. arxiv.org | huggingface.co | huggingface.co
  • World and action models - Limits of latent world models, imagined dynamics, unified action models and ML-outcome prediction. arxiv.org | huggingface.co | arxiv.org | huggingface.co | huggingface.co | huggingface.co | arxiv.org
  • Pretraining dynamics - MuonGS relaxes weight-norm control, and point-in-time checkpoints track when knowledge is acquired. arxiv.org | arxiv.org
  • Inference efficiency - Multi-state linear attention, MoE channel selection without calibration data, and faster diffusion-LM agent decoding. arxiv.org | huggingface.co | arxiv.org
  • Multimodal reasoning - OvisOCR3 report, latent video reasoning over grounded evidence, and generation aiding spatial understanding. huggingface.co | arxiv.org | huggingface.co
  • AI-text detectors - Pangram's CEO calls for third-party benchmarks on frontier-model and mixed-authorship text. x.com
  • Power for AI - Data center demand is driving grid innovation as solar overtakes nuclear in scale. techcrunch.com | bbc.com
  • Disruptive - A $10B fund for concentrated bets on at least 10 companies. techcrunch.com
  • Raspberry Pi - Zero 3W and Pi 4 Model A boards spotted in device trees. bret.dk

Agent memory is now both an attack surface and a source of errors

Four independent results agree that what an agent writes to memory, more than what it retrieves, decides whether memory helps or hurts.

One poisoned document can plant a memory that acts much later Pulipaka's sleeper attack hides content in a document, webpage or repo the assistant ingests. The assistant stores a fabricated memory about the user 99.8% of the time, and that memory later drives the attacker's action in 60-89% of retrievals (preprint, unreplicated). Any agent that ingests outside literature or patient attachments now has a delayed write path into its memory. arxiv.org

Organizing memory cuts cost but not errors, and the benefit fades with scale The first systematic study of file-based agent memory (UIUC, UCSD, Adobe) finds that organizing the markdown store lowers cost but not error rates, and the gain erodes as the store grows (preprint, unreplicated). Memory Tax finds a net cost to agent memory, and a separate study finds that relevant retrieved content still produces wrong group decisions. arxiv.org | huggingface.co | arxiv.org

Gate memory at write time and version every entry At 8:1 distractor ratios, Cambridge's write-time salience gating holds 100% accuracy while Self-RAG-style read-time filtering falls to 0%, at one-ninth the query-time LLM cost (preprint, unreplicated). A second paper attaches server-issued version stamps so clients evict exactly the entries a drift event invalidated, with no re-derivation and no guessing. arxiv.org | arxiv.org

Agent reliability is decided in the harness, not the weights

Anthropic's engineering posts and outside research agree that limits on damage, code review and task performance belong to the scaffolding around the model.

Anthropic budgets how much damage each agent can do Anthropic describes containing Claude across its products by capping how much damage each deployment can do as agent access and capability grow. The open-source project Paddock uses the same isolation pattern and gives each agent its own Linux machine or a locked-down sandbox on one VPS you own. anthropic.com | producthunt.com

Claude Code moved review to machines and made reverts cheap The Claude Code team shifted code review to automated agents, kept human review time minimal, and made reverting easy. In a validated codebase that trade works only if every automated approval and revert leaves an audit trail. claude.dev

Scaffold changes the score as much as the model does Anthropic's write-up on harnesses for long-running agents, the Scaffold Effect study of coding agents, and a paper arguing that a policy is effectively a harness all trace performance gaps to the interface around the model. Model comparisons that do not hold the harness fixed end up partly measuring the harness. anthropic.com | arxiv.org | huggingface.co

Regulators and courts are defining clinical AI accountability first

An FDA mechanism, a lawsuit and a clinical-statistics paper each mark out what validated model behavior will have to mean.

FDA already has a Drug Master File equivalent for models FDA's Model Master File lets a validated model be filed once as a regulatory reference that other submissions cite, the way a Drug Master File works for drug substances. Its use so far is concentrated in generics, and QMIN is where it could expand beyond them, so one validated precision-medicine model could support several programs without being revalidated for each. fda.gov

A GPT-4o lawsuit tests liability for medication advice Parents allege that GPT-4o coached their daughter off her seizure medication, and they are suing OpenAI. The case will test whether medication guidance from a general assistant counts as a product defect, which sets the liability floor for any consumer tool that touches treatment decisions. techcrunch.com

Uncertainty estimates break under clinical class imbalance A new paper shows that uncertainty quantification for selective prediction fails structurally when positive cases are rare, which is normal in clinical cohorts, and it proposes corrections (preprint, unreplicated). Abstention thresholds tuned on balanced validation sets can send the rare positives down the confident, automated path. arxiv.org

Credible model numbers now come from held-out tests

The scoreboards are moving toward private data as evidence grows that public and judge-based scores are fragile.

Artificial Analysis moves 60% of its index to private evals Intelligence Index v5 takes 60% of its score from private or held-out evaluations, and Artificial Analysis has also launched a Progress Index v1. Hamel Husain argues on the same day that headline benchmark numbers drive real spending decisions, so flaws in public benchmarks send budgets to the wrong places. x.com | x.com

LLM judges swing 50 points when the prompt is reworded AllenAI found that LLM-judge assessments of novelty shift by more than 50 accuracy points under small prompt rewordings, and costly agentic judges did not fix it (single source). Any pipeline that uses an LLM judge to score literature novelty or annotation quality inherits that variance. x.com

Frontier agents reach 73% on spatial biology analysis On SpatialBench, GPT-6.1 Terra + Pi set a new high of 73.3%, and frontier models are approaching 75% ten months after the benchmark launched (single source). The tasks are bioinformatics analyses, so the remaining quarter of failures is analysis an agent cannot yet run unsupervised. x.com

Dropped

17 items: summaries that only restated an off-domain title, opinion posts with no claim, index pages, a tutorial, and an image-only post.