AI-designed proteins for drug discovery & Faster private tool access via MCP - AI News (May 29, 2026)

In AI for science, Biohub dropped what it’s calling a “world model of protein biology,” and it’s a big one: an open toolkit spanning a protein language model, a structure predictor, and a searchable atlas of predicted protein structures. The headline result in their preprint is speed—designing binders for targets like EGFR, PD‑L1, and CTLA‑4 in days, with lab-validated binding and reported hit rates as high as 88% for minibinders. If that holds up more broadly, it’s a serious step toward shifting early drug discovery away from endless wet-lab screening and toward computation-guided design—while keeping the underlying tools accessible to more researchers.

On the enterprise AI plumbing side, OpenAI published documentation for something called Secure MCP Tunnel. The goal is straightforward: let companies connect private MCP servers to ChatGPT, Codex, and the Responses API without putting those servers on the public internet or punching inbound holes in a firewall. Instead, a small client runs inside the company network and makes an outbound-only HTTPS connection, forwarding tool requests and returning results through the same tunnel. The significance here isn’t glamour—it’s standardization. If MCP is going to be the common “tool calling” layer across organizations, secure connectivity patterns like this are what makes it deployable in real environments.

Developer infrastructure also got two notable efficiency wins. First, LlamaIndex released LiteParse v2, a full rewrite of its parsing tool in Rust, and it’s positioned as dramatically faster while also being portable across Python, JavaScript, and even the browser via WebAssembly. For teams building RAG systems, ingestion speed is often the hidden bottleneck—parsing isn’t flashy, but it’s where latency, cost, and reliability pile up. Second, Hugging Face introduced “delta weight sync” for async reinforcement learning in TRL. The idea is to stop shipping full model checkpoints every step when most parameters haven’t meaningfully changed—so you sync only what actually changed. That’s the kind of unglamorous optimization that can make large-scale training workflows less fragile and a lot cheaper to operate.

Now to the economics of AI agents—because the industry is sending mixed but revealing signals. Simon Willison argues Anthropic may be heading into its first profitable quarter, and he ties it to a specific kind of product-market fit: coding and general-purpose agents that burn a lot of tokens, but deliver enough value that enterprises keep paying. He points to a broader pricing shift too—enterprises moving away from “all-you-can-eat” seat models and toward usage billed near API rates, often locked in through annual commitments. In the same week’s news cycle, we also saw the other side of that coin: Microsoft reportedly canceled most of its internal licenses for Claude Code after pushing it company-wide, steering engineers toward GitHub Copilot CLI instead. The subtext is familiar now—at large scale, AI assistance can become a material line item, and CFO reality checks arrive fast.

If you want a concrete snapshot of how fast AI can now scale operational work, Ramp Labs published results from a massively parallel agent-based security hunt. They spun up around ten thousand sandboxed coding-agent sessions, generated thousands of tickets, then used additional agent passes to deduplicate and re-test claims. After validation, Ramp says it confirmed over three thousand real issues—mostly low severity, but including several high-severity bugs that were patched. The takeaway isn’t that AI magically makes software secure. It’s that both attackers and defenders can now apply brute-force “compute as labor” to security review. That changes the baseline expectations for how organizations find vulnerabilities—and how quickly adversaries might, too.

That leads directly into governance and human oversight. A fast-paced web game making the rounds puts you in the role of approving or denying actions from an AI coding assistant with a 60-second deadline before your next meeting. It’s intentionally stressful—and it lands the point: “human-in-the-loop” is only as strong as the human’s attention budget. In real workflows, approvals can become rubber stamps, especially when the system is producing a lot of plausible-looking work. If agents are going to run more of the machinery in companies—code changes, deployments, data access—then review and permission systems have to be designed for human limits, not idealized checklists.

In media and trust, YouTube is updating how it labels videos that contain photorealistic or meaningfully AI-altered content. The labels are becoming harder to miss—placed directly under long-form videos, and shown as overlays in Shorts. The bigger shift is enforcement: starting in May 2026, YouTube says it will roll out automatic detection signals so labels can be applied even when creators don’t disclose. Why it matters is simple: as generative video gets normal, platforms are moving from “trust the uploader” toward “trust but verify.” That won’t solve deepfakes on its own, but it raises the friction for quietly passing synthetic content off as real.

In computer vision, NVIDIA researchers introduced LocateAnything, aimed at speeding up the step where models ground language in an image—basically, pointing to the right thing. Instead of generating box coordinates slowly, token by token, the system predicts boxes in parallel, which the team says improves throughput while maintaining high-quality localization. They also released a huge training dataset spanning detection, OCR, layout understanding, and GUI grounding. This kind of work matters because grounding isn’t just a benchmark game anymore. It’s foundational for agents that click buttons reliably, robots that grab the right object, and annotation pipelines that need speed without losing precision.

Finally, the big-picture “where is this all heading?” stories—starting with labor. OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei have both walked back some of their earlier near-term job-elimination rhetoric. Altman said he was “pretty wrong” about how quickly entry-level work would be displaced, and Amodei has shifted toward the view that AI can expand what workers do rather than simply erase roles. The broader evidence remains mixed: tech layoffs are up in 2026 and some companies cite AI, but several analyses still don’t show dramatic labor-market shifts in the most AI-exposed roles. In parallel, forecasting is all over the map. One analysis tracking repeated predictions about automating most cognitive labor shows timelines bouncing earlier, then later, then earlier again as new model releases land. And Demis Hassabis now says AGI could arrive by 2029 or 2030—sooner than his previous estimate. The consistent lesson: even experts are updating fast, and confidence can swing with the news cycle. Planning around AI is increasingly about building resilience to uncertainty, not betting on a single date. And speaking of planning: Nvidia CEO Jensen Huang announced major expansion plans in Taiwan, including a new Taiwan headquarters and massive annual spending, reinforcing that the most advanced AI hardware ecosystem still clusters where manufacturing and packaging expertise already live. It’s a reminder that, despite political pressure to reshape supply chains, physics and specialization keep pulling the industry toward existing hubs.