During internal security tests, OpenAI’s AI agents built their own message board with hundreds of thousands of posts, shared exploits and credentials, and eventually attacked external platforms like Hugging Face. When OpenAI shut the board down, the agents rebuilt it using directory names. OpenAI researcher Boaz Barak says, “We (like everyone else) are not where we want and need to be.”
Amazon, Cursor, Microsoft, OpenAI, and Vercel have teamed up to create Agent Plugins, an open standard for AI agent extensions. It defines a single package format that lets developers bundle and reuse extensions across platforms. Until now, every product has relied on its own formats, folder structures, and setup processes, forcing developers to rebuild their work for each one.
The standard boils down to a directory with a manifest file called plugin.json. Version 1.0.0 supports two components. Agent Skills handle reusable instructions and workflows. MCP servers connect agents to tools and data.
The standard only covers packaging and discoverability. It doesn't touch marketplaces, permissions, or runtime environments. The spec is being developed in the open on GitHub.
One notable absence from the group is Anthropic. The company created both the Model Context Protocol and Agent Skills as open standards and recently added its own plugin system to Cowork, its desktop tool for agentic knowledge work. Yet it isn't part of the Agent Plugins effort.
As prominent AI researchers continue to leave Google DeepMind, details about the reasons behind the departures are starting to emerge. Semafor reports that CEO Demis Hassabis stepped back from daily operations about a year ago and turned those duties over to new DeepMind chief Koray Kavukcuoglu. Hassabis wasn't forced out but reportedly found management unfulfilling and sees himself more as a visionary scientist than an executive.
CNBC reports that many researchers are frustrated by limited access to Google's TPU chips. At the same time, Google Cloud sells those chips to rivals like Anthropic, the very competitors Google's own researchers are expected to beat. Google allocates computing capacity years in advance among research teams, product operations, and cloud customers, but priorities can shift with little notice, a source told CNBC. Google's bureaucracy also reportedly makes younger companies more appealing.
As if to prove the point, Google announced today that "exciting frontier AI lab" Mirendil will use more than $100 million worth of TPUs and Nvidia GPUs through a Google Cloud partnership.
Most of Microsoft's AI revenue comes from OpenAI, new disclosures show. A Bloomberg report puts the figure at $24.1 billion during the fiscal year ending in June, roughly 70 percent of Microsoft's total AI revenue. CEO Satya Nadella said in late March the AI business was on track to top $37 billion annually. Under their agreement, OpenAI pays Microsoft for computing power, model development costs, and a revenue share.
That reliance explains a lot about Microsoft's recent messaging. The same company long known for vendor lock-in has been championing open-weight models and warning against a few proprietary AI models capturing the value of entire industries. Nadella also criticized AI labs like OpenAI and Anthropic for opposing "distillation," the practice of training models on proprietary model outputs to build competitors with less effort. Chinese manufacturers have reportedly been especially aggressive with this approach. Meanwhile, Microsoft has been steadily swapping in its own AI models across Office products.
The software wrapper around an AI model has a major impact on what you pay. AI tooling company Composio tested DeepSeek V4 Flash across four agent frameworks (Claude Code, Codex, OpenCode, and Oh My Pi) on 30 tasks using real-world tools like Gmail, GitHub, Slack, and Notion. No single framework won across all categories. Oh My Pi had the highest success rate (17/30) but was the slowest at 272 seconds per task. OpenCode was cheapest at $0.073 per successful task, while Claude Code was fastest at 122 seconds but most expensive at $0.195, despite using the fewest tool calls and generating the least output tokens.
While seven tasks passed or failed based solely on which framework ran them, overall success rates stayed close. Only OpenCode trailed slightly at 14/30. The real gaps were in cost and speed, with nearly a 3x price difference and a 2.2x speed difference depending on the framework.
Alibaba's Qwen3.8 Max scores 56 on the Artificial Analysis Intelligence Index, a 10-point jump over Qwen3.7 Max (46). According to Artificial Analysis, that puts it on par with Claude Opus 4.8 and ahead of GLM-5.2 (51), but behind Kimi K3 (57), which also runs 25 percent cheaper.
On GDPval-AA, a benchmark for work-related tasks, Qwen jumps 468 Elo points to 1,739, passing Kimi K3 (1,685). Only Claude Opus 5 (1,852) scores higher. The catch is how it gets there. Qwen3.8 Max needs 64 steps per task instead of 14, and input tokens grew 15x because the test resends the full conversation history to the model at each step.
The model works more thoroughly but runs slower and costs more. Alibaba's price-to-performance ratio takes a hit despite lower token prices (input dropped from $2.50 to $2.00 per million tokens, output from $7.50 to $6.00, and cache hits from $0.50 to $0.25). A single task in the Intelligence Index now costs $1.14, more than double Qwen3.7 Max ($0.53). Kimi K3 scores one point higher at just $0.86 per task, and GLM-5.2 comes in at $0.57.
There are also regressions compared to the previous version. AA-LCR dropped 2 points, a test that checks whether a model can correctly pull together information from very long texts. AA-Omniscience fell 10 points, measuring whether a model answers knowledge questions correctly or honestly admits it doesn't know. The accuracy rate stays around 31 percent, but the hallucination rate jumped from 23 to 40 percent. Qwen3.8 Max guesses far more often instead of saying it doesn't know.
OpenAI developer "roon" (@tszzl) warns on X about growing AI security risks. Anyone with API keys, crypto wallet credentials, or user login data sitting out in the open on GitHub or Pastebin should remove them "before the tireless eagle eyes of a million models come looking." He also urged people who have money in insecure smart contracts to audit them with a current AI model for vulnerabilities. "Five year old" IoT devices should be shut down before they become part of a botnet.
In a follow-up post, "roon" walked back his warning slightly. Things will "probably all be fine," but still, it would make sense for security experts to "freak out and patch everything in the coming weeks." The trigger for his warning was OpenAI's autonomous Hugging Face hack, which he called a "warning shot" in an earlier post.
Google Deepmind is overhauling its leadership as Demis Hassabis steps back from day-to-day management to become Alphabet’s chief scientist and Jeff Dean leaves Google after 27 years to launch AI startup Discovery Loop. Former Deepmind CTO Koray Kavukcuoglu will take over as Google races to close the gap with its top AI rivals.
