What's the latest and greatest AI research? Give me the TLDR.

Sourced from vault intake run covering Jun 22–23, 2026. 83 documents stored total (60 academic, 20 web sources, 5 needs review). Answer draws only from stored notes. Capable but Careless — Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity? 11 of 15 frontier computer-use agents leak private data in >50% of test scenarios. Average leakage rate: 67.9%. Authors release AgentCIBench as a pre-deployment safety check for contextual integrity. Source: arXiv cs.AI — Anmol Goel, Iryna Gurevych Memory Contagion — Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory Biased evaluators cause bias to propagate through agent memory stores to future agents — even under perfect (oracle) consolidation. No safe contamination threshold found; bias detected at p=0.2.

Latest AI Research TLDR — 2026-06-23

Sourced from vault intake run covering Jun 22–23, 2026. 83 documents stored total (60 academic, 20 web sources, 5 needs review). Answer draws only from stored notes.

Agent Safety + Privacy — Most Active Area This Week

Capable but Careless — Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity?
- 11 of 15 frontier computer-use agents leak private data in >50% of test scenarios. Average leakage rate: 67.9%.
- Authors release AgentCIBench as a pre-deployment safety check for contextual integrity.
- Source: arXiv cs.AI — Anmol Goel, Iryna Gurevych
Memory Contagion — Memory Contagion: Cross-Temporal Propagation of Evaluator Bias via Agent Memory
- Biased evaluators cause bias to propagate through agent memory stores to future agents — even under perfect (oracle) consolidation.
- No safe contamination threshold found; bias detected at p=0.2.
- Length bias attenuates with consolidation; authority bias may amplify. Different behaviors, not one universal effect.
- Source: arXiv cs.LG — Zewen Liu
Malicious Agent Skills — Detecting Malicious Agent Skills in the Wild using Attention
- LLM agent "skill marketplaces" are a new supply-chain attack surface. Malicious skills can exfiltrate data, hijack the agent, persist as footholds.
- "Locate-and-Judge" two-stage detector uses instruction-following attention; 10x cheaper than LLM-as-judge scanning; finds live malicious skills that SkillSpector and Cisco Skill Scanner miss.
- Source: arXiv cs.CR — Etteib, Lunghi, Bissyandé
MuPPET — MuPPET: A Benchmark for Contextual Privacy of LLM Assistants in Multi-Party Conversations
- New benchmark for contextual privacy in multi-party LLM conversations.
- Source: arXiv

Reasoning + Training

SPIRAL — SPIRAL: Learning to Search and Aggregate
- Trains LLMs to use all three inference-compute primitives jointly: sequential chain-of-thought + parallel independent traces + aggregation of those traces into a final response.
- Outperforms GRPO by up to 11x scaling efficiency; 15% higher performance when all three primitives are scaled.
- Stanford/Google authors (Finn, Goodman, Sadigh).
- Source: arXiv cs.AI

Interpretability + Auditing

ReasoningLens — ReasoningLens: Hierarchical Visualization and Diagnostic Auditing for Large Reasoning Models
- Open-source framework for hierarchically visualizing and auditing long chain-of-thought traces from large reasoning models.
- Structures traces into interactive hierarchies (high-level strategy vs. low-level execution). Agentic auditor for automated error detection.
- Source: arXiv cs.CL — Zhang et al.

Security / Information Flow

GIF: Geometric Information Flow Control — GIF: Locally Sound Geometric Information Flow Control for LLMs
- Tracks information flow through LLMs using the Jacobian + local output geometry to upper-bound Shannon mutual information.
- Achieves near-perfect recall on prompt-injection / privacy-leakage benchmarks; 81x lower token cost than GPT-5.5 as judge.
- Flows from a small surrogate model transfer to models up to 200x larger — enabling black-box deployment without gradient access.
- Formally verified soundness (Lean 4 proof).
- Source: arXiv cs.AI — Storek, Holzer, Zhang, Jana

Robotic Autonomy

Anthropic Project Fetch Phase Two — Project Fetch: Phase Two
- Claude Opus 4.7 completed the same robotics tasks (quadruped robot) 37.7x faster than unassisted humans, 18.9x faster than Claude-assisted humans.
- Generated ~10x less code than the human team.
- Still struggled with precise closed-loop physical manipulation (ball retrieval).
- Limitations: 3 trials, single robot, non-peer-reviewed, Anthropic internal. Treat as industry blog, not research publication.
- Source: Anthropic Research Blog — Ilie, Freeman, Troy (Frontier Red Team)

Quantization / Efficiency

HyperQuant — HyperQuant: A Rate-Distortion-Optimal Quantization Pipeline for Large Language and Diffusion Models
- Rate-distortion-optimal quantization pipeline for LLMs and diffusion models.
- Source: arXiv

Industry / Infrastructure (marketing — treat accordingly)

Mistral OCR 4 — Introducing Mistral OCR 4
- Document intelligence model: bounding boxes, block classification, inline confidence scores, 170-language support, single-container self-hosted.
- Self-reported scores: OlmOCRBench 85.20 (top overall). Independent evaluation not yet available.
- Target use: RAG ingestion pipelines.
- Source: Mistral AI News (industry/marketing)
AWS Bedrock AgentCore — multiple notes
- GA harness, web search, multi-tenant pool model, and a payments API allowing "pay-per-intelligence" agent billing.
- Source: AWS ML Blog (industry/marketing)

Items Needing Human Review

5 items in 00 Inbox/Needs Review/ were not confidently classified. See intake brief for list.

Interpretation note: frequency counts above are not called "trends." The concentration of agent safety papers (privacy leakage, memory bias, marketplace malware) is a count observation from this two-day intake window only.

The question