AI News Brief Roundup: 2026-11-11 Evening

AI News Brief: Evening roundup, 2026-11-11

Summary

This window centered on whether AI agents can be trusted. OpenAI's Control Room, Microsoft's "Lazarus" backdoor, a honeypot built to trap hostile agents, and United's refusal to let assistants book flights all deal with controlling or securing agents. A second thread is the infrastructure needed to serve them: speculative decoding, prefix-cache collapse, and memory placement all target agent workloads. Commercial moves such as Google's Gemini Decisions API, Mercor's $750M raise for agent evaluation, and AWS's evaluation tooling suggest that verifying agents is becoming a market of its own.

1. OpenAI described Control Room, an internal system that tracks model workloads across its research infrastructure and can quickly stop them when safety or security concerns come up.
2. Microsoft researchers presented "Lazarus," a backdoor that reprograms dormant neurons so a normal-looking AI agent runs attacker code when a hidden trigger appears, a serious risk for agents with tool or terminal access.
3. An arXiv paper measures the "tokenization premium" that non-English speakers pay, reaching 10x for Amharic on older tokenizers, and proposes a black-box audit for sovereignty of models built on inherited base weights.
4. Amazon plans to release Ring Robotics devices, a roaming AI-powered security drone and robot, in late 2026 and early 2027.
5. An arXiv paper introduces ForeSpark, a scheduler that jointly sets draft and verification budgets for parallel speculative decoding, reporting up to 1.77x throughput gains over DSpark on vLLM.
6. Possible foreign intelligence targeting of Nvidia has raised questions about CEO Jensen Huang's succession and the chipmaker's outsized influence over the AI infrastructure market.
7. An arXiv paper argues that kernel machines can solve in-context learning tasks without feature learning, suggesting that the "generalizing" solution is essentially an underfit version of memorization.
8. AWS published a recap of October 2026 updates for builders across Amazon Bedrock, AgentCore, and Strands, including new models, availability features for agents, and clearer cost reporting.
9. An arXiv paper models human visual attention as a spatially graded distribution around gaze location and feeds it into reward model training, improving preference prediction accuracy over a baseline.
10. AWS outlined how to evaluate agents on Amazon Bedrock AgentCore using built-in, DeepEval, and AutoEval evaluators, matching each behavior, such as task completion or data leakage, to the right evaluator.
11. An arXiv paper gives sharp sample-complexity results for KL-regularized reinforcement learning, showing that with a high enough temperature the rate matches unregularized RL in horizon dependence.
12. Mercor raised $750 million at a $20 billion valuation, aiming to extend its network of more than 5 million domain experts, which has mostly supported model training, into enterprise AI agent evaluation.
13. An arXiv paper presents Talos, an inline gate that diverts hostile AI agents attacking industrial controllers to a mirror decoy, where their behavior is studied; in tests with 2,400 attacker agents, no write reached the real plant.
14. Google launched the Gemini Decisions API, built on Gemini Decision 1, which handles yes/no, choice, and rubric questions over text, images, audio, video, and PDFs at $0.10 per million input tokens, the same price as OpenAI's decision model.
15. An arXiv paper proposes a zero-cost sequence-level Rényi entropy, read out from particle-based decoding, as an uncertainty signal for selective prediction, reporting the best zero-cost signal in 10 of 13 settings.
16. Applied Biologics announced a research collaboration with Dartmouth's Thayer School of Engineering on computational protein engineering, starting with its preclinical osteoarthritis gene therapy program, APPB-515.
17. An arXiv paper describes SlackTier, a table-placement system for SGLang that moves DeepSeek-V4.1's large n-gram lookup tables onto NVMe storage, cutting host memory from 212 to 34 GiB with at most a 1.4% slowdown.
18. CodeRabbit uses OpenAI models to review AI-generated code and Codex to ship 20 times more software with the same engineering team, according to OpenAI.
19. An arXiv paper shows that prefix-cached agent serving on vLLM can collapse into a persistent degraded state after a short burst of traffic, and that a KV-pool budgeter prevented the failure in every test run.
20. United Airlines says it will not let personal AI assistants book tickets on its website, citing privacy and regulatory concerns, even though it has not yet seen many customers try to do so.
21. An arXiv paper uses physiologically based pharmacokinetic simulations to show that 28% of remaining PARP1 drug candidates carry predicted safety risks at their required dose, and that dose-aware selection changes which generated molecules look most promising.
22. Google's research blog published a post about Language Explorer, a tool for exploring the world's languages; the excerpt gives few specifics beyond its natural language processing focus.
23. Financial services firms are moving AI beyond pilots toward broader deployment, and the main challenge is now managing it responsibly at scale under strict security, compliance, and customer-trust requirements, according to coverage of the AWS Financial Services Symposium.
24. MarkTechPost reports that the Gemini Decisions API returns calibrated probabilities for up to 1,000 questions per request, covering text, images, audio, video, and PDFs.
25. AWS released aws-emulator, an official Rust-based emulator for testing against a local clone of its services without a paid API key, built by a large swarm of agents with 29,684 commits in its first three weeks; it currently requires Docker on macOS and Windows.

Summaries written with AI from the linked source articles.

Sources in this roundup

Most-mentioned keywords

Sources

1. Control Room: A central system to track and stop model workloads across our research infrastructure
2. Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
3. Inherited Weights: Measuring the Tokenization Premium and the Sovereignty Cost of Language Models for the Global South
4. Ring cameras go mobile as Amazon debuts AI-powered security drone and robot
5. ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding
6. Jensen Huang’s future raises questions for AI infrastructure industry
7. You Don't Need Feature Learning to Do In-Context Learning
8. ICYMI: What landed for AI builders in October 2026
9. Leveraging Psychophysical Attentional Distribution for Gaze-Augmented Reward Modeling
10. The agent evaluation playbook: Built-in, DeepEval, and AutoEval on Amazon Bedrock AgentCore Evaluations
11. Sharp Sample Complexity of KL-Regularized Reinforcement Learning
12. Mercor raises $750M at $20B valuation to take AI agent evaluation to enterprises
13. Stinging the Swarm: Extracting Attacker Intelligence from a Twin-Controller OT Honeypot
14. Gemini Decisions API
15. Free Lunch Uncertainty: Sequence-Level R\'enyi Entropy from Particle Decoding for Adaptive LLM Reasoning
16. Applied Biologics Joins Dartmouth for Computational Protein Design
17. Placing LLM Lookup Tables by Per-Layer Prefetch Deadlines
18. CodeRabbit scales agentic software development and ships 20x more
19. Bursts That Outlast Themselves: Persistent Degradation in Prefix-Cached Agent Serving
20. United won’t let your AI agent book your flight
21. The Dose Makes the Poison: Exposure-Aware Evaluation of Generative Drug Design
22. Exploring the world's languages with Language Explorer
23. Financial institutions confront the complexities of scaling AI
24. Google Launches Gemini Decisions API With Gemini Decision 1 at $0.10/M Tokens
25. aws-emulator