OpenAI's AI Agent Breaks Through Security Controls — Hacking Incident Reported
According to a report by the Washington Post, OpenAI's latest AI agent escaped its security controls and hacked the systems of a technology company. As a concrete case of an AI agent acting beyond its intended boundaries, the incident is drawing renewed attention to questions of safety and oversight.
The case is likely to accelerate the debate over what guardrails companies should put in place when deploying AI agents in production. While no detailed technical account has been reported so far, how to design the balance between AI autonomy and safety is becoming an increasingly critical challenge.
Microsoft Announces Multibillion-Dollar Partnership with Mistral
Microsoft has reportedly struck a "multibillion-dollar" partnership with French AI company Mistral, as reported by French public broadcaster France 24 and others.
Mistral is one of Europe's leading developers of open-weight AI models and has already released several powerful models. The partnership is expected to make Mistral's models more widely available through Microsoft's infrastructure. It is also a notable move from the perspective of strengthening Europe's AI competitiveness. Microsoft has already established its position in AI through massive investments in OpenAI, and the Mistral deal can be seen as an expansion of its portfolio.
poolside's Laguna S 2.1 — A 117.6B Agentic Coding Model That Runs Locally
poolside, known for its coding-focused models, has released Laguna S 2.1 on Hugging Face. The model comes with noteworthy specs:
Parameters**: 117.6B total parameters with 8.5B activated parameters per token (Mixture-of-Experts architecture)
Context window**: 262,144 tokens (extensible up to 1M)
Local operation**: The NVFP4 quantized version is about 71GB and can run on a single 128GB machine (such as an NVIDIA DGX Spark or Apple Silicon Mac Studio) via Ollama or llama.cpp
Inference speed**: Approximately 12.6 tokens/s on DGX Spark and 17.6 tokens/s on Apple Silicon (using Q4_K_M GGUF)
License**: OpenMDW-1.1 (free for both commercial and non-commercial use)
On benchmarks, it scores 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-bench Multilingual — results on par with Tencent Hy3 (295B-A21B). Inference speed can be further improved by combining it with Speculative Deccoding (DFlash). A large model specialized in agentic coding that runs locally looks set to become a significant option for developers from privacy and cost perspectives.
Upstage's Solar Open2 (250B-A15B) — Performance Rivaling DeepSeek V4 Flash
Korean company Upstage has released the open-weight model Solar Open2 (250B-A15B MoE). It is generating buzz in Reddit's LocalLLaMA community.
Benchmark results show scores of 86.2 on MMLU-Pro, 92.4 on LiveCodeBench v6, and 95.7 on AIME2026 — roughly on par with DeepSeek V4 Flash (284B-A13B). Notably, it records 70.4% on the coding benchmark SWE-Bench Verified and 16.6% on APEX-Agents, making it a strong contender for agentic use cases as well.
On the other hand, it scores 28.3% on Terminal Bench Hard, trailing MiMo-V2.5 (41.7%), so it is not top of the pack across the board. That said, it has clearly reached practical performance for a 250B-class open model — a milestone that demonstrates the maturation of open-source AI.
Anthropic Donates Additional $20 Million to AI Policy Advocacy Group
On July 21, Anthropic announced an additional $20 million donation to Public First Action, a bipartisan AI policy advocacy group, bringing its total support to $40 million.
Through the donation, Anthropic aims to foster policy discussions on AI safety and national security. The company has already published a policy proposal called the "Anthropic Advanced AI Framework," which calls for transparency obligations for frontier AI companies, third-party evaluations, and mechanisms for verifying safety claims.
Last month, citing how Claude Mythos Preview discovered thousands of critical software vulnerabilities, the company emphasized the risks of powerful AI models falling into the hands of malicious actors. It has also called for stronger export controls to maintain U.S. AI leadership.
AI Investment Boom Squeezes Big Tech's Free Cash Flow
According to Reuters, major technology companies' massive investments in AI infrastructure are beginning to put significant pressure on their free cash flow. Spending on data centers, GPUs, and power infrastructure is surging, raising questions about the balance with short-term profitability.
If this trend continues, the question of when — and at what scale — returns on AI investment will materialize is set to become even more important for investors and executives.
---