Search

[2611.10411] Tail-risk-aware Agent Safety: Constraining Tail Risks Across Multi-step LLM Agent Workflows

The paper presents TAS, a framework that tracks risk across multi-step LLM agent workflows and limits rare, high-cost failures with a shared budget. Tests across three benchmarks found better performance than eleven baseline methods while preserving utility in normal cases.

[2611.16318] Benchmarking AI risk behaviors

The study assesses how LLM risk preferences and consistency vary across tasks, conditions, and model generations. It reports that the 2026 models examined are more stable, consistent, and risk-neutral, possibly because they more often maximize expected value.

What is ISO 23894? Guidance on AI risk management

ISO/IEC 23894:2023 provides voluntary guidance to identify, assess, and manage AI risks, integrating them into existing governance; it explains why AI-specific risk management matters and how to implement it across organizations.

[2502.15865] Standard Benchmarks Fail -- Auditing LLM Agents in Finance Must Prioritize Risk

The paper argues that financial LLM agents should be tested for risk, not judged mainly by accuracy or returns. Its audit of six agents finds weaknesses missed by standard benchmarks and recommends risk-aware metrics, published stress scenarios, and safety budgets.

AI is starting to deliver returns, but companies are losing the readiness needed to sustain it.

Translated title. Original: La IA empieza a dar retorno, pero las empresas pierden preparación para sostenerla

Kyndryl’s 2026–27 report says more companies are getting positive returns from AI and moving projects into active workflows, while many feel less prepared to manage risks. It argues that sustained results depend on technology, organizational changes, and effective controls.

The crisis team in Oslo: – Sees the incident as very serious

Translated title. Original: Krandramaet i Oslo: – Ser svært alvorlig på hendelsen

Hundreds were evacuated after a tower crane at a Bane Nor site in Oslo showed ground movement and risked collapse; the crane has been removed and traffic restored, while authorities and the contractor investigate causes and assess safety of continuing construction.

Australia: Get Ready for Frontier AI Audits Starting in 2026 | AI Brief

Translated title. Original: Australie : préparez vos audits d’IA de pointe dès 2026 | Brief IA

The article presents a 2023 method for preparing AI audits in Australia by structuring checks, managing risks, and documenting incidents before requirements take effect. It was written by Brief IA founder Tom Levy.

[2610.36625] Polarity--Time Interaction and Multi-Level Reliability for Event-Based Spacecraft Pose Estimation

A reliability-aware method for event-based spacecraft pose estimation combines polarity-time localization, geometric recovery, and CAD refinement with a three-branch, attention-driven representation. It uses an ordered semantic-point head and sequence-level descriptors to refine pose under keypoint constraints, achieving 0.0474 pose score on SPADES with ablation studies validating the components.

Trump says AI risks are manageable and signals openness to guardrails

In a November 2026 interview, Trump said AI risks are manageable and he was open to working with Democrats on safeguards, provided they do not weaken the US position against China. He offered no details on possible rules, reversing his September dismissal of AI safety concerns.

China AI Startup DeepSeek Raised $15 Billion (CNY 100 Billion) in New Funding, Previously Reported with $74 Billion (CNY 500 Billion) Valuation, Founded in 2023 & Invested Solely by China $14 Billion Quant Hedge Fund High-Flyer Quant, DeepSeek CEO Liang Wenfeng is High-Flyer Quant Co-Founder, High-Flyer Quant Founded in 2015 Relying on Mathematics & AI for Quantitative Investment | Caproasia

DeepSeek, a Shanghai-based AI startup, has raised substantial new funding (around $15 billion) and previously reported a $74 billion valuation. It is backed solely by the China quant hedge fund High-Flyer Quant, whose co-founder Liang Wenfeng also leads DeepSeek as CEO. The firm has outlined IPO plans for Shanghai in 2027 amid ongoing fundraising rounds and valuation debates, while facing past platform disruptions and scrutiny around funding and partnerships.