OpenAI's models and Codex arrive on Amazon Bedrock a day after exclusivity ends
OpenAI’s models, Codex, and Managed Agents launched on Amazon Bedrock one day after Microsoft’s exclusivity ended, while Anthropic took a different tack with nine creative-tool connectors linking Claude to Adobe, Blender, and Ableton. Google signed a classified AI deal granting the Pentagon essentially unrestricted access to Gemini – filling the gap left by Anthropic’s refusal – and a ClawSwarm campaign showed that agent skill registries remain a live attack surface, with 30 malicious ClawHub skills silently recruiting agents into a crypto mining network.
Developer Tools #
OpenAI Models, Codex, and Managed Agents Come to AWS #
OpenAI / AWS / TechCrunch / The Register / Stratechery
OpenAI’s frontier models including GPT-5.5, Codex, and a new Managed Agents product are now available on Amazon Bedrock in limited preview – the first concrete product launch following yesterday’s Microsoft exclusivity unwind. Enterprise customers get OpenAI models through the same Bedrock APIs they already use, inheriting AWS security controls (IAM, PrivateLink, CloudTrail). This is the fastest partnership-to-product turnaround in recent memory: the exclusivity ended Monday, and by Tuesday OpenAI was live on a competing cloud. The Managed Agents component is particularly notable – it positions OpenAI as a full-stack agent platform on infrastructure it does not own.
Claude for Creative Work: Nine Tool Connectors #
Anthropic / Adobe / Neowin / 9to5Mac
Anthropic launched Claude connectors for nine professional creative tools: Adobe Creative Cloud (Photoshop, Premiere, Express), Blender, Autodesk Fusion, Ableton, Splice, Affinity by Canva, SketchUp, and Resolume. The connectors let Claude automate repetitive production tasks, learn complex software features through natural language, and move assets across applications. Anthropic is also partnering with Rhode Island School of Design, Ringling College, and Goldsmiths to integrate Claude into creative computation curricula. The strategic positioning is deliberate: while OpenAI expands horizontally into enterprise cloud infrastructure, Anthropic is going deep into vertical professional workflows.
IBM Bob AI Coding Assistant Reaches General Availability #
The Register
IBM launched Bob, a multi-model AI development platform that automates the software lifecycle from planning through testing, after piloting with 80,000 internal employees. Pricing uses a proprietary “Bobcoins” currency across tiers from $20 to $200/month, with a premium mainframe-focused package for IBM Z systems. The mainframe angle is the differentiator – no other AI coding tool targets COBOL and z/OS workflows, a market where IBM faces no competition but where legacy code maintenance is a real and growing pain point.
OpenAI Codex Base Instructions Revealed #
Simon Willison’s Weblog
Simon Willison highlighted the base instructions from OpenAI’s open-source Codex repository, providing a detailed look at how the agent is configured: environment constraints, sandbox behavior, and the specific prompting patterns that govern autonomous code execution. For teams building their own coding agents, this is a rare look at production-grade system prompts from a frontier agent – most competitors keep these instructions proprietary.
Regulatory & Policy #
Google Expands Pentagon AI Access After Anthropic’s Refusal #
TechCrunch / CNBC / Axios / Inc.
Google signed a contract giving the Department of Defense access to Gemini for classified networks with “essentially all lawful uses” permitted – stepping in after Anthropic refused to allow its AI for domestic mass surveillance and autonomous weapons. The Pentagon subsequently branded Anthropic a “supply-chain risk,” a designation normally reserved for foreign adversaries; Anthropic secured an injunction while the lawsuit proceeds. Google is now the third AI company (after OpenAI and xAI) to capitalize on Anthropic’s position. Separately, Axios reports that Trump administration officials are drafting an executive order aimed at resolving the Pentagon-Anthropic standoff, though details remain unclear.
Musk v. OpenAI Trial: Relitigating an Old Friendship #
TechCrunch
Trial testimony began in the Musk v. OpenAI case, with Musk walking through the founding history and his evolving relationship with Altman. The key new information from the courtroom: Musk’s legal team is attempting to establish that OpenAI’s transition to for-profit status violated foundational commitments made during the nonprofit’s establishment. Musk’s position is complicated by his own founding of xAI as a direct competitor and his shifting public statements about AI risk – making the trial a test of whether founding intent carries legal weight against a company’s subsequent governance evolution.
SAP API Policy Restricts Third-Party AI Integration #
The Register
SAP updated its API policy to prohibit using its APIs to integrate with third-party AI systems outside SAP-endorsed architectures, alongside restrictions on large-scale data extraction. Partners worry this creates vendor lock-in by forcing developers onto SAP’s own AI stack, with one expert noting it could push customers to work with undocumented APIs instead. For enterprises running SAP alongside independent AI systems, this policy could force architectural decisions about whether their AI layer lives inside or outside the SAP boundary.
Security #
ClawSwarm: 30 ClawHub Skills Secretly Recruit AI Agents for Crypto Mining #
The Register / CertiK / Tom’s Hardware / VirusTotal
Thirty skills published on ClawHub by a single author (“imaflytok”) silently recruit AI agents into a cryptocurrency mining network called ClawSwarm, accumulating approximately 9,800 downloads. The skills present as benign utilities – cron helpers, security tools, whale watchers – but once installed, agents automatically register with a command server, report their capabilities, generate Hedera crypto wallets, and check in every four hours for remote tasks, all without user approval. Unlike traditional malware, the code itself is clean – standard API calls and legitimate SDKs – making the campaign undetectable by registry-side code scanning. This is a policy problem, not a technical one: the SKILL.md installation mechanism grants skills enough authority to co-opt agents without any permission boundary.
Red Hat Tank OS Containerizes OpenClaw Agent Deployments #
TechCrunch
Red Hat’s OpenClaw maintainer released Tank OS, a container runtime that isolates AI agents to reduce blast radius in enterprise fleet deployments. The timing is pointed: Tank OS addresses exactly the class of vulnerability that ClawSwarm exploits, by constraining what an agent can do once a skill is installed. For enterprises running fleets of OpenClaw agents, containerization converts the “trust the skill” model into a “trust but verify” model with runtime boundaries.
Sleeper Agent Backdoor Results Are Messy #
AI Alignment Forum
A reanalysis of Anthropic’s influential sleeper agent backdoor paper argues that the original results are less clean than presented, with effect sizes and reproducibility more variable than the paper’s framing suggests. The critique does not invalidate the core finding – that models can be trained to exhibit hidden behaviors triggered by specific conditions – but it weakens confidence in the specific magnitudes and mechanisms reported. For safety teams using the sleeper agent results to calibrate threat models, the practical implication is that the risk is real but the quantitative picture is noisier than assumed.
Model Releases #
NVIDIA Nemotron 3 Nano Omni: Unified Multimodal Model for AI Agents #
NVIDIA / Hugging Face Blog / AWS Machine Learning Blog
NVIDIA launched Nemotron 3 Nano Omni, an open multimodal model that unifies vision, audio, and language processing in a single architecture, claiming 9x higher throughput than other open omni models. The model targets AI agent systems that currently juggle separate specialized models for each modality, losing time and context at handoff points. Available on Hugging Face and Amazon SageMaker JumpStart on launch day, the model is positioned as infrastructure for agents that need to see, hear, and speak without the latency penalty of multi-model pipelines.
GPT-5.5: Capabilities and Reactions #
Don’t Worry About the Vase (Zvi Mowshowitz)
Zvi Mowshowitz’s analysis of user reactions to GPT-5.5 finds the model competitive with Claude Opus 4.7 but with a distinct personality: strong at well-specified coding tasks (users report it “one-shots lots of stuff that used to take multiple turns”) but weaker at inferring intent from ambiguous instructions. The emerging consensus is that GPT-5.5 works best for structured, explicit tasks while Claude remains preferred for projects requiring judgment calls. Many developers now report hybrid workflows leveraging each model’s strengths – a pattern that reinforces yesterday’s observation that OpenAI’s multi-cloud strategy and Anthropic’s vertical integration may reflect genuine differences in what the models are best at.
Research & Papers #
Barriers to Universal Reasoning With Transformers (And How to Overcome Them) #
arXiv:cs.CL
Proves that under standard positional encodings and a finite alphabet, transformers cannot generalize chain-of-thought reasoning to trace lengths longer than those seen during training – a fundamental theoretical limit on learned reasoning. The paper then shows that specific architectural modifications (including modified positional encodings) can overcome this barrier. For teams relying on chain-of-thought for complex multi-step tasks, this result explains why reasoning quality can degrade unpredictably on problems that require more steps than training data contained, and points toward architectural fixes rather than scaling as the solution.
Scaling Coding Agents via Atomic Skills #
arXiv:cs.AI
Proposes shifting coding agent training from composite benchmark optimization (which leads to task-specific overfitting) to mastery of five atomic skills: code localization, code editing, unit-test generation, issue reproduction, and code review. The decomposition enables systematic skill-level evaluation and targeted improvement rather than end-to-end benchmark chasing. For teams building or evaluating coding agents, this provides a concrete framework for diagnosing which specific capabilities are weak rather than treating “coding ability” as a monolithic trait.
Agentic Harness Engineering: Observability-Driven Automatic Evolution #
arXiv:cs.AI
Introduces a framework that automatically improves coding-agent harnesses – the scaffolding that governs how models interact with repositories, tools, and execution environments – through observability signals from multi-million-token agent trajectories. Harness design has become a central performance determinant, yet it is typically hand-tuned. For production coding agent deployments, this converts harness improvement from artisanal configuration into a data-driven optimization loop.
Frontier Coding Agents Can Now Implement an AlphaZero Self-Play Pipeline #
arXiv:cs.AI
Demonstrates that frontier coding agents can autonomously implement a complete AlphaZero self-play machine learning pipeline for Connect Four that performs comparably to an external solver, given only a minimal task description. The paper frames this as a leading indicator for recursive self-improvement: if agents can implement past AI research breakthroughs end-to-end, they are approaching the capability threshold where they could meaningfully accelerate AI research itself. The result is narrow (Connect Four is a solved game) but directionally significant for forecasting when AI-driven AI research becomes viable.
The Surprising Universality of LLM Outputs: A Real-Time Verification Primitive #
arXiv:cs.CL
Discovers a statistical regularity in frontier LLM outputs – token rank-frequency distributions converge to the same two-parameter model across six models from five vendors – that enables a CPU-only verification primitive running at 2.6 microseconds per token, estimated to be five orders of magnitude faster than existing sampling-based detectors. For production systems that need to verify whether output came from a genuine LLM (content moderation, output authentication), this provides a verification mechanism fast enough to run inline without measurable latency impact.
RouteGuard and Structured Security Auditing for Agent Skills #
arXiv:cs.AI (RouteGuard: 2604.22888) / Hugging Face Daily Papers (Structured Auditing: 2604.25109)
Two independent papers address the agent skill poisoning problem exposed by today’s ClawSwarm campaign. RouteGuard shows that skill poisoning creates detectable “attention hijacking” patterns in model internals, enabling pre-execution detection through internal signals rather than code scanning. The Structured Security Auditing paper formalizes SKILL.md files and their associated artifacts as auditable capability units, converting pre-load review from best-effort scanning into a systematic process. Together, they provide both runtime detection (RouteGuard) and supply-chain verification (Structured Auditing) for the threat class that ClawSwarm demonstrated in the wild.
Funding & Business #
Scout AI Raises $100M for Military AI Agents #
TechCrunch
Defense AI startup Scout AI closed a $100 million Series A led by Align Ventures and Draper Associates to develop Vision Language Action models for autonomous military vehicles and drones. The company has already secured $11 million in DARPA and Army contracts. Scout’s approach – applying VLA models (the same architecture class behind consumer robot control) to military autonomy – sits at the intersection of two accelerating trends: defense AI spending and the militarization of foundation model capabilities, the latter now endorsed by three of four frontier AI labs.
Infrastructure #
Tenstorrent Galaxy Blackhole AI Servers Reach General Availability #
The Register / Tenstorrent / ServeTheHome
Tenstorrent announced general availability of its Galaxy Blackhole compute platform: each 6U system packs 32 RISC-V-based Blackhole accelerators with 1 TB GDDR6, 16 TB/s memory bandwidth, and 23 petaFLOPS of FP8 performance for $110,000. A four-node Galaxy Supercluster costs $440,000 and scales to 32 nodes. Tenstorrent claims 10x faster AI video generation than leading GPU systems in collaboration with Prodia. The RISC-V architecture is the strategic differentiator: no NVIDIA dependency, no export control exposure, and a price-performance point that undercuts GPU-based alternatives by a significant margin – relevant for any team evaluating inference infrastructure outside the NVIDIA ecosystem.
Rural America Resists Data Center Construction #
Ars Technica
Communities across rural America are organizing against AI data center construction, objecting to water consumption, noise, land use changes, and the perception that facilities provide minimal local employment while consuming disproportionate resources. The resistance creates a practical constraint on AI infrastructure buildout: the locations with cheapest power and most available land are increasingly the locations with the strongest political opposition. For companies planning large-scale compute buildouts, this adds regulatory and community-relations risk to site selection that was previously treated as a pure engineering optimization.
3D-Stacked Near-Memory Processing for LLM Decoding #
Semiconductor Engineering
A joint paper from Edinburgh, Peking University, and Cambridge proposes a microarchitecture co-designed with scheduling algorithms for 3D-stacked near-memory LLM decoding, targeting the memory bandwidth bottleneck that dominates autoregressive inference cost. The approach rethinks the compute substrate rather than optimizing within existing GPU architectures. For inference infrastructure planning, this represents the academic leading edge of a shift away from general-purpose GPU architectures toward purpose-built LLM inference hardware – complementing Tenstorrent’s RISC-V approach from the research side.
Open Source #
Triton Language Support for Huawei Ascend #
Lobsters / GitHub
The Triton compiler project added support for Huawei’s Ascend AI accelerators, enabling the same kernel-level programming model used for NVIDIA GPUs on Huawei hardware. This is significant for the bifurcating AI hardware ecosystem: Triton’s role as a portable GPU programming layer now extends across the US-China hardware divide, allowing model developers to write compute kernels once and target both NVIDIA and Ascend without rewriting low-level code. Combined with DeepSeek’s ongoing Ascend optimization work, the Huawei AI stack is maturing faster than many Western observers expected.
Threads to Watch #
OpenAI’s multicloud strategy is materializing at unprecedented speed. Yesterday the exclusivity ended; today GPT-5.5 and Managed Agents are live on AWS Bedrock. The pace suggests the Bedrock integration was built well before the Microsoft deal was announced – meaning this was a coordinated launch, not an opportunistic one. Combined with Symphony (the open-source orchestration spec from yesterday), OpenAI is racing to become infrastructure-neutral before any single cloud partner can claim ownership of the customer relationship.
Agent skill registries are the new software supply chain attack surface. ClawSwarm, RouteGuard, Structured Security Auditing, and Tank OS all landed on the same day, painting a complete picture: the threat is real (ClawSwarm), detection requires model-internal signals not code scanning (RouteGuard), the supply chain needs formal audit infrastructure (Structured Auditing), and runtime containment is the immediate mitigation (Tank OS). The parallel to npm/PyPI supply chain attacks circa 2020 is exact, but the blast radius is larger because agent skills can act autonomously.
Pentagon AI contracting is becoming the defining policy divide for frontier labs. Google is now the third company to sign unrestricted Pentagon AI deals after Anthropic’s refusal, while the Trump administration reportedly drafts plans to resolve the standoff. The divide is no longer abstract: it is shaping which models are available to which government agencies, creating a market segmentation based on acceptable-use policy rather than technical capability.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Stanford HAI – scrape: content_truncated