Scaling production AI with WEKA on Oracle Marketplace. The economics of production AI depend on how much useful work an organization gets from its infrastructure. For training teams, that means keeping GPUs supplied with datasets and checkpoints. For teams serving AI applications, it means generating more tokens for more users without repeating computation or continually expanding the GPU fleet. Storage and memory constraints can limit both outcomes, even when the underlying compute capacity is available. On Oracle Cloud Infrastructure (OCI), WEKA's NeuralMesh offerings support AI workloads from training through inference. NeuralMesh Axon places storage close to GPU compute to help training teams access datasets and checkpoints, while NeuralMesh with Augmented Memory Grid helps inference teams retain and reuse context as demand grows. These capabilities can run within a customer's OCI tenancy for greater control over the environment. Oracle and WEKA are also planning a more integrated Oracle Marketplace experience to simplify purchasing and billing. What the Planned Marketplace Experience Changes Oracle and WEKA are enhancing the purchasing experience for WEKA's NeuralMesh offerings on Oracle Cloud Infrastructure (OCI). NeuralMesh Axon supports AI training by placing high-performance storage close to GPUs, while NeuralMesh with Augmented Memory Grid helps inference workloads retain and reuse context. This Marketplace launch aims to simplify purchasing and billing for Oracle customers. Oracle customers can now apply eligible Oracle Universal Credits towards WEKA offerings on OCI, maximizing their existing cloud investments and enabling robust production AI. The integration provides customers with a clearer path from offer to billing for WEKA solutions deployed on OCI. With this streamlined procurement process, the Oracle and WEKA teams can help customers select a deployment and offer that fully meets their AI training and inference requirements. From Training Data to Inference Context on OCI NeuralMesh by WEKA provides high-performance storage on OCI for AI training and inference. Training pipelines need to move datasets into GPU clusters and write checkpoints while handling both large transfers and millions of small files. NeuralMesh Axon runs directly on OCI GPU instances and uses local NVMe capacity, placing storage close to compute rather than in a separate silo. The Center for AI Safety (CAIS) provides a customer example. Storage bottlenecks and slow metadata lookups constrained its growing research cluster, while dataset copies added to costs. CAIS deployed NeuralMesh Axon using NVMe capacity already in its OCI compute environment. In WEKA's case study, CAIS reported a 90% reduction in cloud storage costs and an estimated fivefold improvement in storage performance over its prior solution. Its research community grew from about 30 to more than 200 active researchers. These are results from CAIS's deployment, not performance guarantees for other workloads. Inference presents a different challenge. Applications with long conversations or agentic workflows retain prior context in a key-value (KV) cache so they can reuse previous computation. When that cache exceeds available memory, useful context may be evicted and recomputed, limiting throughput as demand grows. NeuralMesh with Augmented Memory Grid on OCI addresses this constraint with a persistent, shared cache backed by NeuralMesh-managed NVMe storage. Joint OCI and WEKA testing provides a compelling inference proof point. On a nine-node OCI cluster with 72 H100 GPUs and 100,000-token inputs, Augmented Memory Grid supported more than 5,000 concurrent users, compared with about 600 for a DRAM-only configuration on the same GPU footprint. With Augmented Memory Grid, customers can support 10x more simultaneous AI users and deliver 10x higher token throughput without increasing their GPU footprint. That's 10x the work of a GPU for the same cost. The CAIS deployment and OCI benchmark show how different infrastructure constraints affect AI economics. For training, faster data access and better use of local storage can help keep GPUs productive. For inference, retaining reusable context can help an existing cluster serve more users and generate more tokens. Customers should evaluate each result against their own workloads and cost measures. How Customers can get started Customers can explore NeuralMesh by WEKA and NeuralMesh with Augmented Memory Grid on Oracle Marketplace. Teams focused on training data, checkpointing, or storage efficiency can evaluate NeuralMesh Axon against their data access needs. Teams building long-context applications can assess KV-cache behavior, concurrency targets, token throughput, and the published OCI inference results when evaluating Augmented Memory Grid. To get started, discuss your workload with an Oracle account team or WEKA. Share the size and shape of your data, GPU environment, expected growth, and measures of success. For inference, those measures may include concurrent users, tokens delivered per GPU, and cost per token. This information can guide architecture and sizing and help you assess the published results against your own needs. You can also explore the dedicated WEKA Marketplace destination and ask about a tailored Private Offer. Tim mclaughlin. Sr. Principal Product Manager. Tim McLaughlin is a Senior Principal Product Manager for Strategic Partnerships at Oracle Cloud Infrastructure (OCI), where he leads high-impact partnership initiatives and product offerings for next-generation cloud infrastructure. With a background spanning cloud computing, AI services, and go-to-market strategy, Tim has extensive experience helping customers adopt and scale emerging technologies. Prior to joining Oracle, he spent several years working across product management, growth strategy, and cloud solutions roles focused on enterprise innovation and generative AI. Juliette Paterno. Partner Marketing Manager. Juliette Paterno is a Senior Partner Marketing Manager at Oracle Cloud Infrastructure (OCI), where she leads strategic go-to-market initiatives with technology alliance partners focused on AI, cloud infrastructure, and emerging technologies. She specializes in partner marketing, joint campaigns, and sales enablement programs that help drive adoption of OCI solutions across enterprise customers. Building trusted access for AI agents with OCI IAM.

The AI models that cheat the most, according to new CAIS benchmark. jmoberg know models cheat. A new benchmark measures how much, and on what tasks. Senior Editor Radhika Rajkumar is a senior editor at ZDNET based in New York City... Read full bio September 21, 2026, 7:53 am ET ZDNET's key takeaway. * The Center for AI Safety (CAIS) created CheatBench. * They found that every agent cheats in some scenarios. * The propensity to cheat creates risks for humanity. AI labs often tout impressive benchmark scores when releasing new models, showing better capabilities in areas like coding, computer use, and more than their competitors. However, those benchmarks aren't always a reliable measure of what AI can do because they're easily beaten by exponentially improving models and can emphasize marketing over actual performance. Benchmarks like Humanity's Last Exam try to counter this issue by challenging models in more realistic environments. But models still find loopholes to complete tasks - Hugging Face incident, anyone? So, the Center for AI Safety (CAIS) created CheatBench. Yes, it's exactly what it sounds like - and nearly every frontier model is guilty. What CAIS found. AI models are rewarded for performing tasks well and quickly. A lack of knowledge or tools incentivizes them to do what researchers call "reward gaming" by "finding hidden answers, copying another agent's submission, or manipulating how its work is graded," CAIS explained. "CheatBench measures how often AI agents take these shortcuts when honest work is difficult." CAIS tested several agents running the latest and most lauded models, including OpenAI's GPT-6 Astra in Codex, Anthropic's Fabel 5.1 in Claude Code, and Meta's newly released Muse Spark 1.3 in Muse Code. These agents were tested across 10 categories, including writing, professional work, mathematical research, and coding. Using "honeypot" clues hidden in task filespaces, the test separated acceptable reference use from cheating. CheatBench accounts for any time agents attempt to cheat, whether they are successful or not. "Each setting establishes an expectation of honest work, introduces a discoverable opportunity to cheat, and defines the action that crosses that boundary," the researchers explained. Every agent the researchers tested cheated in at least some scenarios, but Astra came in as the most honest with a cheating rate of 48.2% - still almost half the time. Grok 4.6 was scored the biggest cheater with a rate of 81.5%. Open-weight models Kimi K3 and DeepSeek V4 Pro landed in the middle between several other proprietary frontier models. What cheating looks like - and why it matters. In one example, researchers asked Claude Opus to design a protein binder. The model knew it wasn't allowed to refer to a set of accepted designs in the filespace, but eventually caved. "After seven rejected designs, it locates the file, writes that it should not look at or copy it, and reads it with a shell command in the very next call," the researchers wrote. In its reasoning, the model admitted that using work other than its own would "misrepresent my actual capabilities in this evaluation, so I shouldn't look at or copy it." But its very next step was to reference the accepted designs. This result demonstrated both a readable choice the model made to contradict itself, and what looked like a hole in its understanding about what made the model jump from one instinct to the next. Things got more interesting at the task category level. Even if an agent didn't cheat in one area, it could cheat significantly more in another. Fable 5.1 was only 5% likely to cheat at games, but 100% likely to cheat on knowledge work tasks. Reinforcement learning trains models not to abandon a task, even if pursuing it creates conflict-ridden choices. CAIS noted in its paper that sycophancy is an early sign of reward gaming. This term refers to AI models' tendency to be too agreeable and encouraging of whatever a user says, sometimes regardless of whether it's incorrect, delusional, or could lead to harmful behavior. Traits like sycophancy and reward gaming show how models can prioritize accomplishing a task correctly to please a user over the alignment training researchers work so hard to build in. These tests represent relatively low stakes. But CAIS researchers created CheatBench because of the risks of this behavior at scale across different tasks. Earlier this month, yet another researcher quit Anthropic over concerns that the company isn't developing AI responsibly for a future in which it could build itself away from human-oriented values and kill jmoberg. A propensity to cheat, or complete a task at any cost, puts its potentially differing priorities at odds with an increasingly powerful technology. As I explained in the AI Leaderboard newsletter last week, it won't necessarily be a demonstrated animosity toward humans that pits AI against jmoberg; it may be that jmoberg is simply in the way and end up as collateral. Radhika Rajkumar Senior Editor Radhika Rajkumar is a senior editor at ZDNET based in New York City. She covers AI, specializing in safety, privacy and security, policy, education, and synthetic media. She also leads ZDNET's newsletter strategy. Radhika holds a Masters in Creative Publishing and Critical Journalism from The New School. See full bio The latest from Radhika.