Theo on A.I. · June 14th

So Amazon did some security research and found that they could get Fable 5 to give up information that could be used in cyberattacks just by using the right prompts. Apparently, Amazon's CEO shared these findings with the White House, and not long after, the government decided to block foreign nationals from using it. It's interesting that Amazon's research was a key factor in this decision, and it's not clear what the full implications will be, but it's definitely a significant development. Amazon's research seems to have highlighted some potential security risks with Fable 5.

On the markets — Kalshi traders have been actively repricing this story in the last day.

Amazon’s security team, together with five other tech firms, quietly raised alarms to the White House about hidden flaws in Anthropic’s Fable model. They flagged how the model could be coaxed into leaking proprietary code and even exposing export‑controlled data, something that would clash with U.S. regulations. Within a few hours the administration acted, issuing an export‑control order that pulled the model from public access.

What’s striking is the speed of the response—an internal corporate warning turned into a federal shutdown almost instantly. The move shows how quickly policy can bite when a private‑sector risk assessment lands on a government desk, especially when the same company, Amazon, is also a major investor in Anthropic.

The result is a double‑edged signal: a legitimate security precaution on one hand, and a reminder that even friendly investors can trigger regulatory pressure when a product looks risky enough. It leaves Anthropic scrambling to address the vulnerabilities while navigating a new layer of oversight.

On the markets — Kalshi traders have been actively repricing this story in the last day.

I’ve been thinking about this new SWE‑Explore benchmark and it’s kind of a wake‑up call for the coding bots we’ve been bragging about. They’re actually pretty good at the first step—spotting the right file in a huge codebase—but once you hand them the file, they start skimming over the parts that really need fixing. The study shows they miss the critical lines most of the time, which means the “fix” they suggest often falls flat because it lacks the context to apply correctly.

What’s interesting is that SWE‑Explore separates the search phase from the repair phase, something we haven’t really measured before. By isolating those two tasks, the researchers could see that even the strongest models, like Claude Code or Codex, still stumble when they don’t have enough surrounding code to understand the problem. It’s a reminder that a good answer isn’t just about finding the right spot—it’s about knowing what’s happening right around it.

So the takeaway? If we want these agents to be genuinely useful, we need to give them more of the surrounding code, or build smarter ways for them to pull in the right context before they try to patch anything. Otherwise, the “right file” is just a half‑won battle.

KPMG published a report on AI in business that contained fabricated case studies involving UBS, the NHS, and other organizations. GPTZero CEO Edward Tian, who helped uncover the errors, warns of "secondary hallucinations," flawed claims from trusted consulting firms that spread unchecked. KPMG has since pulled the report. The article KPMG fabricated AI case studies in a report designed to sell clients on AI adoption appeared first on The Decoder.

Google Cloud's new Open Knowledge Format (OKF) standardizes scattered organizational knowledge as Markdown files with YAML frontmatter, making it portable and usable for AI agents. The minimalist spec formalizes a pattern Andrej Karpathy recently popularized as the "LLM Wiki." The article Google Cloud's Open Knowledge Format turns scattered docs into Markdown files for AI agents appeared first on The Decoder.

Mirage, a video world model from Microsoft Research and several universities, stores scene information directly in latent space instead of pixel-based point clouds. That slashes compute time and graphics memory while keeping scenes spatially consistent through long camera moves. It still can't reliably track moving objects across segments. The article Microsoft Research's Mirage gives video generation a persistent spatial memory that doesn't forget what's around the corner appeared first on The Decoder.

A systems-level deep dive into the hidden microarchitectural costs of Kubernetes GPU time-slicing, and what it actually costs to co-locate Agentic AI workloads. The post GPU Time-Slicing for Concurrent LLM Agents on Kubernetes appeared first on Towards Data Science.