One-Line Summary

Government and enterprise channels for AI both tighten and expand: OpenAI opens compliant routes for agencies and AWS customers while briefing Congress on cyber risks, China blocks Meta’s Manus deal, Citi lifts AI’s 2030 market to $4.2T, and new agentic tools land for creatives and ops.

Big Tech

OpenAI and Anthropic brief House staff on cyber-capable models

OpenAI and Anthropic met with staff of the House Homeland Security Committee in classified briefings to explain their new cyber‑capable AI models and what they could mean for critical infrastructure. Axios reports that Anthropic withholds a public release of its Mythos Preview due to its ability to quickly find and exploit critical flaws, and that OpenAI is using a tiered approach for its GPT‑5.4‑Cyber model while working to give federal agencies access. The briefings also touched on alleged industrial‑scale attempts to copy U.S. AI models. ¹

For government adoption, OpenAI says ChatGPT Enterprise and its API Platform now hold FedRAMP 20x Moderate authorization, enabling U.S. federal agencies to use managed products — including access to GPT‑5.5 in OpenAI’s FedRAMP environment — subject to each agency’s decisions. OpenAI highlights cloud‑native security evidence, Key Security Indicators, and reusable authorization data for review in its Trust Portal. ²

OpenAI is also expanding distribution to enterprise IT by bringing its models — including GPT‑5.5 — to Amazon Bedrock, plus offering Codex on AWS and launching Bedrock Managed Agents powered by OpenAI in limited preview. The company frames this as a way for organizations to build within existing AWS security, identity, and procurement workflows. ³

In a separate statement of intent, OpenAI published updated operating principles emphasizing democratization, resilience, and iterative deployment, and says it will collaborate with governments and other actors to address risks — including those arising from increasingly capable models — while adapting its stance as it learns more. ⁴

China blocks Meta’s Manus acquisition amid AI rivalry

China’s top economic planning agency prohibits the foreign acquisition of Manus and orders the parties to withdraw from the deal, halting Meta’s planned purchase of the Singapore‑based AI agent startup over security review concerns. The decision follows an earlier probe and arrives as Meta seeks to expand agent capabilities across its platforms. ⁵

Analysts quoted in regional media cast the block as a warning that tech, talent, and data links to China can trigger intervention even when a startup has re‑incorporated overseas, suggesting a new pressure point in U.S.–China AI competition. ⁶

Coverage also notes the move signals tighter scrutiny of foreign participation in sensitive AI sectors, complicating cross‑border deals and creating additional uncertainty for U.S. firms attempting to acquire China‑rooted AI assets. ⁷

Industry & Biz

Citi lifts AI market forecast to $4.2T on enterprise demand

Citigroup raises its global AI market outlook to more than $4.2 trillion by 2030, with roughly $1.9 trillion tied to enterprise AI, citing faster‑than‑expected adoption of coding and automation tools and strong revenue growth at providers like Anthropic. The prior estimate was more than $3.5 trillion overall, with nearly $1.2 trillion from enterprise AI. ⁸

Citi’s note characterizes Anthropic as “the leader in enterprise AI,” pointing to traction in software development and agentic, task‑automating workflows; it also cites a business mix with about 80% of revenue from enterprises and large compute‑capacity deals — up to $40 billion from Google and as much as $25 billion from Amazon. ⁹

The brokerage says Anthropic’s annualized revenue run rate surpasses $30 billion by April, while competition from OpenAI, Google, and others is pushing the battle toward workflow integration and reliability rather than pure model benchmarks. ¹⁰

New Tools

Adobe’s Firefly AI Assistant enters public beta

Adobe’s Firefly AI Assistant is a chat‑style creative helper inside Adobe Firefly that executes multi‑step workflows across Photoshop, Lightroom, Premiere, and more; it is now in public beta. Adobe positions the tool to turn plain‑language requests — like turning a product shot into a full set of social assets or building a mood board — into finished outputs while keeping creators in control. ¹¹

The assistant can draw from 60+ pro‑grade tools (for example, Generative Fill, Auto Tone, Remove Background, Vectorize, and Presets) and includes Creative Skills — pre‑built workflows for common jobs like batch photo edits, portrait retouching, social variations, and product mockups. ¹²

During the beta, availability rolls out globally for customers on Creative Cloud Pro or paid Firefly plans (Pro, Pro Plus, Premium), with complimentary daily generative credits for use with the assistant. ¹³

DigitalOcean debuts Inference Engine for cheaper, scalable AI

DigitalOcean launches an Inference Engine that combines four capabilities — an Inference Router, Batch Inference, Serverless Inference, and Dedicated Inference — so teams can match workloads to performance and cost profiles, with customers reporting up to 67% lower inference costs. The company says the goal is unified control over how production inference runs and scales. ¹⁴

According to the announcement, the Inference Router uses a mixture‑of‑experts (MoE) router model to send each request to the right model based on task and developer priorities; Batch Inference targets up to a 50% cost reduction for offline jobs; Serverless adds scale‑to‑zero and off‑peak pricing; and Dedicated Inference provides reserved capacity. DigitalOcean cites results from Artificial Analysis showing 3× faster time‑to‑first‑token and 3× higher output speed than Amazon Bedrock on one DeepSeek test at 10,000 input tokens. ¹⁵

Customer case studies highlight gains such as 2× production throughput with 40% lower P99 latency at Hippocratic AI and up to 67% lower costs at Workato’s research lab, alongside faster time‑to‑first‑token and lower end‑to‑end latency. ¹⁶

What This Means for You

If you work with U.S. public‑sector customers or heavily regulated industries, OpenAI’s FedRAMP Moderate status for ChatGPT Enterprise and its API removes a major procurement and compliance barrier — including access to GPT‑5.5 inside the FedRAMP environment — which can speed pilots for drafting, translation, and knowledge work. Bring this into security reviews and RFP responses immediately. ²

For enterprise IT already on AWS, OpenAI models on Amazon Bedrock — plus Codex and Bedrock Managed Agents powered by OpenAI — reduce friction by fitting into existing identity, security, and billing. This positions AI projects to move from trials to production within the tools and controls your teams already use. ³

The classified briefings underscore that lawmakers view cyber‑capable models as a near‑term risk. Product, marketing, design, and ops leaders should coordinate with SecOps on stronger approval flows for external tools, and refresh training on social‑engineering and content‑authenticity defenses as models get better at offensive tasks. ¹

Budgets are likely to shift toward enterprise AI that automates workflows end‑to‑end. Citi’s upgrade to a $4.2 trillion market — with $1.9 trillion tied to enterprise AI — signals executives are prioritizing coding assistants, agentic workflows, and integrated copilots; frame your Q2 experiments in terms of cycle‑time cuts and measurable ROI. ⁸

Action Items

1. **Try Adobe Firefly AI Assistant ** (public beta): If you have Creative Cloud Pro or a paid Firefly plan, use one product photo to generate platform‑specific social assets and time how long it takes versus your current process.
2. Pilot OpenAI on AWS Bedrock with IT: Ask your AWS admin to enable Bedrock access for a small team and test one internal workflow (e.g., requirements summarization) using an OpenAI model within your existing security and billing setup.
3. Leverage FedRAMP in public‑sector sales: If you sell to U.S. agencies or contractors, update security questionnaires and proposal templates to note ChatGPT Enterprise and the OpenAI API are available at FedRAMP Moderate, including access to GPT‑5.5.
4. Run a 30‑minute phishing and deepfake drill: With your security lead, review two recent scam examples and rehearse a verification checklist (second‑channel confirmation, media provenance) to reduce business‑email‑compromise risk.