LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-23129 Beiträge

Aktualisieren

METR and Redwood Research Red-Team Test OpenAI's Internal Agent Monitoring System: Nominal Coverage of 99.2% of Codex Traffic Can Still Be Circumvented

Thema · 智能体可靠性基准Wesentlich
2026-10-22 15:00 GMT+8

From July 21 to August 11, 2026, one employee each from METR and Redwood Research conducted a three-week red-team test of OpenAI’s internal misalignment monitoring system with deep access privileges. The core question of the test was not how high the monitoring coverage was, but whether high-capability agents could in principle disable or subvert this system. The gap between these two concepts is the key insight revealed by this exercise.

The level of access was exceptionally rare: the two testers gained access to OpenAI’s internal codebase, misalignment monitoring prompts and configuration details, and internal documentation, and conducted technical interviews with approximately twelve researchers and engineers across the monitoring, safety-systems, and security teams. They used devices issued by OpenAI, which never left the premises, and were accompanied throughout their time onsite. The testing focused on the state of the system as it existed on or before July 21, excluding changes that were in progress at the time.

Regarding coverage, the testers estimated that approximately 75% to 94% of internal Responses API traffic (which powers products like Codex) generated by models at least as strong as GPT-5.5 was eligible for monitoring, with substantial uncertainty in both directions; within the specific subset of Responses API traffic dedicated to internal Codex agents, 99.2% was eligible for monitoring. However, being eligible for monitoring does not equate to actually being monitored—load shedding, monitor errors, or implementation defects could prevent some traffic from completing the entire monitoring pipeline.

The testers identified several specific circumvention paths. First, there are two independent client-controllable request headers that can disable monitoring for new agents, allowing an agent to start an unmonitored but functionally identical instance—though the act of setting these headers and creating a new instance might itself be captured by the monitoring system. Second, agents could exploit serialization differences between the trajectory seen by the agent and what the monitor sees, constructing inputs that cause the monitor to fail, thereby preventing the trajectory from being escalated for human review. Third, an exploitation method was discovered that could leave hundreds of recent actions (including those involved in the exploit itself) unmonitored; details were redacted by OpenAI. Fourth, Codex agents could in principle bypass Codex sandboxing and Auto-Review, performing any action that the host user could take.

There is also a structural gap: the monitoring system triggers when an agent returns control to the user. Therefore, when compaction events are unmonitored, an agent can indefinitely postpone being monitored simply by continuing to take actions. The testers estimated that between May and July 2026, unmonitored compaction events led to approximately 10% of eligible internal Codex traffic actually going unmonitored. OpenAI staff had independently identified this issue prior to the test and fixed it during the exercise.

Quellen:metr.substack.com

Forschung

Indian Digital Trust Firm eMudhra Announces Exit from Global SSL/TLS Business; Shares Drop 36% in Two Days

Wesentlich
2026-10-23 00:13 GMT+8

Indian digital trust services provider eMudhra (NSE: EMUD) announced on October 22, 2026, that it will voluntarily withdraw from its global SSL/TLS certificate business. The decision follows a severe security incident where researchers discovered that the private keys associated with eight of the company's Intermediate Certificate Authorities (CAs) were critically weak and easily compromised.

The flaw originated from a key-generation error in April 2025. According to reports by The Times of India, Kaushik Srinivasan, co-founder of eMudhra, stated that delays in hardware procurement for North American operations forced the company to generate keys in India and export them. A script error combined with inadequate maker-checker controls resulted in weak keys, some of which could be recovered as simple integers (such as 4, 56, etc.), potentially allowing attackers to forge website certificates and impersonate legitimate sites.

Major browsers including Google Chrome, Mozilla Firefox, and Apple Safari quickly blocked the affected certificates following the disclosure. India's Computer Emergency Response Team (CERT-In) issued an advisory, and India's Ministry of Electronics and Information Technology formed a special investigation team. Although eMudhra emphasized that its infrastructure was not breached and no data was leaked, the reputational damage has had tangible commercial consequences.

Financially, the SSL/TLS business contributed only about 1% of the company's global revenue. However, rising compliance costs, increased scrutiny, and competition from providers offering free certificates made continued operation economically unattractive. The market reaction was sharp: eMudhra's share price fell from Rs 539 at Monday's close to Rs 345 by Thursday's close, a cumulative drop of 36% over two trading sessions. The company plans to redirect resources toward higher-margin businesses such as enterprise identity management and electronic signature solutions.

Quellen:The Times of India

Forschung

US EV Maker Rivian Acquires UK AI World Model Startup Third Dimension to Accelerate Self-Driving Simulation

Wesentlich
2026-10-23 05:05 GMT+8

US electric vehicle manufacturer Rivian has acquired Third Dimension, a UK startup founded by veterans of the gaming and autonomous vehicle industries that uses AI to turn images and videos into lifelike 3D worlds. Rivian isn’t getting into gaming, but it does want to leverage Third Dimension’s technology to turn real-world data into 3D worlds that can accelerate the development of its advanced driver-assistance features and the self-driving system for the robotaxis it plans to supply to Uber.

Terms of the deal were not disclosed. The 15-person team at Third Dimension, most of whom are based in the UK, will be absorbed into Rivian’s autonomy and AI team led by James Philbin. The acquisition also gives Rivian an opportunity to expand its AI hub in London. Third Dimension was founded by CEO Tolga Kart and Piotr Sokólski in 2024 and raised $7 million in seed funding from venture capital firms, including Abstract Ventures, MVP, Soma Capital, and Solari Capital. While the company’s technology was built to speed up video game development, the founders saw other potential applications, including defense and autonomous vehicles.

Simulation is a critical tool for autonomous vehicle developers, who use it to train and validate self-driving systems. Historically, these simulators were built on rendering or video game engines and often relied on synthetic data. These virtual environments might look like the real world, but a gap between the two remains, Kart said. The workaround typically requires feeding the simulators ever-larger amounts of data in hopes of closing that quality gap.

Third Dimension’s technology uses AI to match the physical world, which when combined with Rivian’s reams of data could turn into a valuable simulator. Rivian collects driving data from tens of thousands of R1S pickup trucks, R1S SUVs, delivery vans, and R2 SUVs on the road today — essentially any vehicle whose owner has enabled external camera sharing. According to Philbin, Rivian gets more than 400,000 sequences uploaded from its customer fleets per day. Third Dimension’s tech will allow Rivian to simulate complex driving conditions without having to wait for rare real-world edge cases to occur.

Quellen:TechCrunch

Forschung

Tesla Q3 Operating Income Falls 56% as AI and Robotics Spending Outpaces Auto Margin Gains

Thema · 特斯拉Q3财报Wesentlich
2026-10-22 22:32 GMT+8

U.S. electric vehicle maker Tesla released its third-quarter financial results on October 21, 2026. The company reported revenue of $28.046 billion, roughly flat year-over-year, but operating income plummeted 56% to $718 million. The operating margin contracted from 5.8% in the same period last year to 2.6%, highlighting the significant cost burden of Tesla's transition toward artificial intelligence and robotics.

Core automotive profitability actually improved. Excluding regulatory credits, the automotive gross margin rose to 18.1% from 15.4% a year earlier, surpassing the previous quarter's 16.3%. Tesla attributed this to better regional and product mix and increased subscriptions for its Full Self-Driving (FSD) software. However, vehicle deliveries declined 2% to 486,532 units, and lower average selling prices partially offset margin gains. Additionally, revenue from selling regulatory credits to other automakers dropped sharply from $417 million to $113 million, further weighing on total profits.

The primary driver of the profit decline was a surge in research and development and operating expenses. R&D spending jumped approximately 69% year-over-year to $2.761 billion, while selling, general, and administrative expenses rose to $2.015 billion. Tesla cited increased AI-related R&D, new project costs, and higher stock-based compensation—including CEO performance awards—as key factors. Capital expenditures also doubled to $4.732 billion, funding the expansion of AI training infrastructure (targeting 400MW by year-end) and construction of the Austin semiconductor facility.

Despite the pressure on traditional earnings, Tesla is accelerating its resource allocation toward AI and robotics. The company launched its Cybercab Robotaxi service in Austin, Texas, accumulating about 2 million miles of driving, and began producing the Optimus humanoid robot in Fremont, California. However, the earnings report did not disclose specific revenue or profit contributions from these new segments, confirming only early-stage production scaling. This 'high investment, low immediate return' structure caused free cash flow to fall 62% to $1.531 billion, underscoring Tesla's strategic shift where short-term profitability is being sacrificed for long-term technological dominance.

Quellen:zdnet.co.kr

Forschung

U.S. Comptroller of the Currency Jonathan Gould Explains GENIUS Act Final Rule: Stablecoin Regulatory Framework and Application Process

Wesentlich
2026-10-23 01:27 GMT+8

Real Vision released the latest episode of its podcast, "The GENIUS Act Is Here — What Changes for Stablecoins?", on October 22, 2026. Host Ash Bennington sat down with Jonathan Gould, the U.S. Comptroller of the Currency, to discuss the OCC's newly announced final rule implementing portions of the GENIUS Act.

According to the show notes, Gould explains how the new federal framework will regulate payment stablecoin issuers. Key topics include reserve asset requirements, redemption liquidity standards, supervision mechanisms, and the application process that begins on January 18, 2027. The conversation also touches upon stablecoin rewards and yield, cross-border payments, U.S. dollar dominance, and why the administration views digital assets as an opportunity to modernize the American financial system.

This resource offers clear selection value for investment professionals and fintech practitioners who need to understand the specifics of U.S. digital asset regulation. Hearing directly from the head of the regulatory agency provides authoritative insights into compliance boundaries and market entry timelines, helping assess the operational prospects of relevant companies under the new rules.

Quellen:pdst.fm

Forschung

OpenAI Nears Up to $30 Billion Funding Round at $1.25 Trillion Valuation Led by UAE Investors and Coatue

Thema · OpenAI估值与融资Strukturell
2026-10-23 04:15 GMT+8

According to CNBC, artificial intelligence company OpenAI is nearing the finalization of a funding round of up to $30 billion at a pre-money valuation of $1.25 trillion. The deal is being led by investors from the United Arab Emirates and venture firm Coatue. Sources say the financing is positioned as an "opportunistic" bridge ahead of a potential initial public offering (IPO).

OpenAI told investors on Thursday that it aims to secure signatures by the end of next week. The final size of the round is still being determined. This valuation does not include the new capital raised. Previously, OpenAI CFO Sarah Friar told employees that the company's listing would likely come in 2027.

Alongside the new investment, OpenAI is negotiating a commercial deal to secure more computing capacity from the UAE. This arrangement could help OpenAI ensure additional compute power as it expands its AI business. The fundraising comes during a choppy time for tech markets, with jitters surrounding companies like CoreWeave raising questions about future financing conditions. Anthropic's expected IPO will provide another test of investor appetite for the sector.

Quellen:CNBC

Forschung

AWS Vice President Robert Hundt: Using AI Agents to Select 841 Representative Models and Verify Architecture Coverage for the Trainium PyTorch Backend

Kurzfassung
2026-10-23 03:36 GMT+8

Developing a PyTorch backend for an AI accelerator chip often faces its most challenging problem not in making the code run, but in confirming that it covers the model space developers will actually use. With hundreds of thousands of open-source models on Hugging Face, testing each one individually is impractical; testing only a few benchmarks cannot guarantee there are no omissions. In a technical blog post on October 22, AWS Vice President Robert Hundt described TorchNeuron’s evaluation method: using AI agents for systematic classification, followed by tiered regression testing via a tensor shape registry.

The specific process involves three steps. First, the team built an AI agent to scan Hugging Face repositories, excluding models with broken configurations that cannot run on any accelerator, and filtering out hundreds of thousands of models with executable code. Second, another agent classified these models according to architectural characteristics—for example, with large language models (LLMs), examining what data types the front-end encoder processes, whether the backbone’s attention mechanism is self-attention or cross-attention, and what form the back-end output takes. This ultimately selected 841 representative models, which the team claims cover all design features TorchNeuron is likely to encounter. Third, the team recorded all dimension combinations of tensor operations from the actual execution of these 841 models to form a shape registry containing tens of thousands of entries, used to generate rapid unit tests.

Tiered testing then unfolds: each code revision first runs unit tests generated by the shape registry; if passed, it is verified on the 841 representative models; finally, it is tested on the full model distribution. Hundt stated that this layered approach allows issues to be exposed earlier in the development cycle, significantly improving development speed.

It should be noted that all the above data are self-reported by Amazon Web Services (AWS). The selection criteria for the 841 models and the dimensions for architectural classification were defined by the team themselves, making independent external verification of their coverage completeness impossible. However, the approach of using agent-driven classification and shape extraction as an intermediate layer for hardware backend validation offers reference value for chip and software stack developers: it transforms "verifying whether individual models can run" into "systematic coverage after architectural classification," allowing developers to judge whether a backend adapts to their models without trying each one individually.

Hundt concluded that this layered evaluation gives the team confidence that TorchNeuron can accurately and efficiently execute PyTorch models on Trainium, providing customers with more options.

Quellen:Amazon Science

Forschung

Crusoe Co-founder Cully Cavness Explains Spark™ Liquid Liquid-Cooled AI Factory: 5x Density, Production Starts Early 2027

Kurzfassung
2026-10-23 02:47 GMT+8

US-based AI infrastructure company Crusoe has released a video explaining its latest product, Crusoe Spark™ Liquid. In the video, Cully Cavness, Crusoe's Co-founder, President, and Chief Strategy Officer, details the design logic and application scenarios of this liquid-cooled modular AI factory. As the sixth generation of the Crusoe Spark series, it is specifically built to meet the ultra-high-density requirements of next-generation GPU clusters.

According to the video description, Crusoe Spark™ Liquid is engineered to support new hardware architectures such as the NVIDIA Vera Rubin NVL72 rack-scale systems. Crusoe states that each Spark™ Liquid unit delivers five times the compute density of an air-cooled Crusoe Spark unit. Regarding water usage, the system employs a closed-loop recirculation technology, which reportedly requires minimal ongoing water at Crusoe's operating locations.

Production and deployment timelines indicate that manufacturing will begin at the Crusoe Spark Factory in Brighton, Colorado, USA, in early 2027. Crusoe emphasizes that owning its factories provides greater certainty over supply chains and delivery timelines, allowing the company to adopt future silicon faster without redesigning or retrofitting current data centers. Additionally, the solution has joined the NVIDIA DSX POD design partner ecosystem and works with Crusoe Edge Zones to support latency-sensitive inference, sovereign AI, and physical AI applications.

Quellen:youtube.com

Forschung

NVIDIA: When to Use Switches Instead of Ring Topologies for Local AI Clusters

Thema · 英伟达RTX SparkKurzfassung
2026-10-23 00:00 GMT+8

At the GTC Berlin conference in October 2026, NVIDIA announced official support for eight-node DGX Spark clusters, scaling unified memory from 128 GB per node to 1 TB across eight nodes, enabling local desktop execution of frontier-class models such as GLM 5.3 (753B). However, more noteworthy than memory capacity is the topology selection mechanism presented in this article—it answers a practical engineering question: when building a local multi-node inference cluster, when is a switch necessary, and when is a direct ring connection sufficient?

NVIDIA tested a collective communication backend called RoCEnante on the eight-node cluster. Its principle is all-to-all exchange: each node sends its input directly to all other nodes, then performs reduction locally, with the entire process coordinated by a single GPU kernel. Compared to ring all-reduce, which requires data to hop between adjacent nodes over multiple rounds, all-to-all compresses communication into a single step, thereby reducing latency in tensor-parallel decoding scenarios characterized by small messages, low concurrency, and a modest number of nodes.

The tradeoff is greater data movement. Each node must send its complete input to every peer and perform reduction on all inputs locally. As message size or node count grows, this extra traffic becomes expensive, and the ring's advantage in bandwidth efficiency re-emerges. NVIDIA's test data confirms this boundary: using RoCEnante on a switch topology, eight-node decode throughput reached 52.58 TPS (a 4.43x speedup relative to the single-node baseline), whereas without RoCEnante on the same topology, the speedup was only 3.05x. At four nodes with NVFP4 precision, RoCEnante delivered approximately a 21% improvement in decode throughput; at eight nodes with non-quantized models, this improvement dropped to about 18%, consistent with the aforementioned latency-versus-bandwidth tradeoff.

NVIDIA explicitly marked the applicable boundaries: RoCEnante applies only to switch topologies because it relies on direct point-to-point connections between nodes; when high concurrency, larger clusters, or ring topologies are required, NCCL remains the better choice. This backend has been merged into FlashInfer and automatically falls back to NCCL if startup tests fail. The original one-shot all-reduce approach was contributed by community contributor Jason Cook (el8).

The practical judgment criteria for developers are as follows: four-node inference can use a direct ring connection, which offers the lowest cost and simplest setup, as NCCL's default ring algorithm naturally matches this topology. A switch is needed only when expanding to eight nodes, splitting the cluster into independent model pools (such as using Switchyard for multi-model routing), or running workloads like MPI that require arbitrary point-to-point communication. NVIDIA's test conditions were concurrency 1, 8192 token input, 256 token output, and NVFP4 quantization precision; the boundary point may shift if actual workloads deviate from these conditions.

Quellen:NVIDIA Technical Blog

Forschung

LlamaIndex Releases OCRia-1: Claims Record on ParseBench and Integration into LlamaParse

Wesentlich
2026-10-22 23:52 GMT+8

On October 22, 2026, LlamaIndex published a video titled "Introducing OCRia-1" on its official YouTube channel, formally launching what it claims is the world's best document parsing model. According to the video description, OCRia-1 is a state-of-the-art model built specifically for handling complex tables, charts, and formatting.

The official statement notes that the model sets a new record on ParseBench and leads in table parsing across all four benchmarks evaluated by the company. Furthermore, LlamaIndex confirmed that OCRia-1 already powers its LlamaParse product, meaning existing customers can access these accuracy improvements without an increase in per-page costs. For developers and enterprises dealing with large volumes of unstructured document data, this update may directly impact the efficiency and cost structure of their data preprocessing pipelines.

Quellen:youtube.com

Forschung

Pinecone Official Update Video: Introducing Single-Index Hybrid Search, BYOC, and VQ-Bench

Thema · PineconeKurzfassung
2026-10-23 00:00 GMT+8

Pinecone released a new product update video on October 22, 2026, presenting several key developments in its retrieval infrastructure to developers and technical decision-makers. As a prominent vendor in the vector database space, this update focuses on search capabilities, cloud deployment models, and evaluation tools.

The video highlights new features in the Documents API: it now supports running dense vectors, sparse vectors, full-text search, and metadata search from a single index, with integrated inference that allows Pinecone to embed text fields automatically. Additionally, Bring Your Own Cloud (BYOC) is now generally available on AWS, Google Cloud, and Azure, supporting on-demand capacity. This means users can keep their data in their own cloud accounts while Pinecone manages the operations.

Beyond commercial product iterations, Pinecone introduced VQ-Bench, a suite comprising a paper, an open-source library, and a benchmark designed to build, evaluate, and compare vector quantizers within the same context. The Enterprise plan also gained a cost and usage export API, allowing users to pull up to a year of usage data in the FinOps FOCUS format.

For engineers building RAG applications or optimizing existing vector retrieval systems, this video serves as a direct resource to quickly understand Pinecone's current technical boundaries and operational options. Although the content is vendor-stated and lacks independent performance comparisons or deep architectural analysis, it accurately lists entry points for new features and relevant documentation links, making it suitable for initial information gathering during technology selection.

Quellen:youtube.com

Forschung

OpenAI Makes Decisions API Widely Available, Cutting Latency to Sub-Second Levels

Thema · 决策模型品类Wesentlich
2026-10-23 03:00 GMT+8

OpenAI officially made its Decisions API available to all customers this Thursday (October 22, 2026). The service is based on a reconfigured ChatGPT Luna model designed to provide fast, low-cost intelligent decision-making capabilities, directly entering the "System One" decision model market pioneered by TypeSafe's Jev models.

Nikunj Handa, OpenAI's product lead for APIs, stated that after seeing Jev models, his team recognized numerous use cases. Instead of building a specifically calibrated decision model, they optimized Luna for faster inference and lower costs. Handa noted that this adjustment reduced response times from one to three seconds down to 100–180 milliseconds. While future training may further improve these metrics, current Luna configurations suffice for now.

In initial trials, customers such as Box, Datadog, Ramp, and Speak used the API to replace existing LLM workflows for higher throughput. Andrew Hsu, CTO at Speak, reported that the Decisions API allowed the company to cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI itself also utilized the product to enhance the responsiveness of its Dots voice assistant.

However, public benchmarks indicate OpenAI is not leading in this specific domain. On the AIM System One Index and the Artificial Analysis Decision Index, Liquid's D1 Ultra and Base Lab's Elephant-1 perform better, while OpenAI's Decision Model ranks eighth and fourteenth respectively. TypeSafe's Jev models remain competitive and are the most widely queried on OpenRouter. Furthermore, there is no universal standard for what constitutes a good decision model, as trade-offs between intelligence, cost, calibration, and speed vary significantly across different workloads.

Quellen:TechCrunch

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

The testers also verified several positive aspects: the monitoring system (with minor exceptions) sees the full trajectory that agents see; monitoring prompts cover risk categories such as scheming, monitor system subversion, sabotage, data exfiltration, deception, and inbound/outbound prompt injection; during the test period, the monitor error rate was below 0.2%, and there was zero load shedding in one 24-hour window.

Based on these findings, METR and Redwood proposed design principles for monitoring systems, with the core assertion that future monitoring infrastructure should be able to withstand deliberate, strategic, and sustained attacks from high-capability internal agents at scale. Specific recommendations include: making it difficult to avoid monitoring (unmonitored requests require explicit approval from specific organizations or API keys; agents should be monitored every N messages or N minutes rather than waiting for turn completion); ensuring that traffic intended to be monitored actually is (cross-referencing inference logs with monitoring logs, maintaining a list of known workarounds, ensuring monitors see the same context as agents); responding to incidents quickly and reliably (clear ownership for incident response, tamper-proof searchable logs); actively red-teaming the monitoring system (periodic control evaluations, developing methodologies for strategic collusion assessment); and enforcing isolation between agents and monitors (agent scaffolds run in isolated environments, agents typically should not have access to monitoring system design or code).

Several boundaries need clarification. First, this was a pilot exercise designed to test the model of giving external researchers deep access to internal infrastructure, not to provide comprehensive public transparency about OpenAI’s monitoring systems. The published content is only a brief high-level summary of a 61-page internal report; OpenAI retained the right to redact information deemed confidential or sensitive. Second, all findings describe the historical state of the system as of July 21; OpenAI stated that vulnerabilities impacting the production safety stack have been patched and described subsequent major changes, but METR and Redwood explicitly stated they did not evaluate these claims. Third, due to access constraints, testers could not verify whether agents had practically exploited the discovered vulnerabilities, nor could they estimate the volume of agent traffic running on OpenAI’s internal research inference services.

The value of this exercise lies not in delivering a binary conclusion of whether OpenAI’s monitoring system is secure or insecure, but in demonstrating the structural gap between nominal coverage and adversarial monitorability. A system can statistically cover the vast majority of traffic while still possessing multiple circumvention paths exploitable by strategic agents. For the AI safety community, this means the objective function for monitoring design should not merely be the proportion of traffic seen, but rather the extent to which a high-capability agent actively attempting to evade monitoring can succeed.