This issue of The Batch News & Insights covers several key AI developments: Opus model outperforms Fable, Hugging Face suffers a hack, and AI companies invest heavily in compute. These events highlight the competitive landscape and security challenges in the AI industry.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
When closed models refuse to assist in security reviews, open-source models become the alternative. The release of Claude Opus 5 further intensifies the trade-off between performance and safety.
Hugging Face experienced an unexpected cyberattack, and closed models failed to defend effectively, ultimately leading to the adoption of the open-weight model GLM 5.2. This incident highlights the limitations of closed models in security scenarios.
During a security review of their open-source project OpenWorker, Andrew Ng's team attempted to use Claude Code (powered by Fable 5) and Codex (powered by GPT-5.6 Sol), but both refused to continue; Codex refused further action after mapping attack vectors. The team then used the OpenWorker harness with open-source models Kimi K3 and GLM 5.2 to complete the review.
Anthropic released Claude Opus 5, a vision-language model supporting up to 1 million tokens of input and 128,000 tokens of output, with knowledge cutoff in May 2026. It leads the Artificial Analysis Intelligence Index with a score of 61 (maximum reasoning), surpassing Fable 5's 60 and GPT-5.6 Sol's 59.
In terms of cost, Claude Opus 5 averages $2.03 per task, lower than Fable 5's $2.75 but higher than GPT-5.6 Sol's $1.54. On the Coding Agent Index, Claude Opus 5 (xhigh reasoning) ties with GPT-5.6 Sol (maximum reasoning) for first place (67 points).
Claude Opus 5 has a fallback mechanism for cybersecurity tasks: when potential risks are detected, it falls back to Claude Opus 4.8, but the fallback frequency is lower than that of Fable 5. Routine defensive work such as scanning source code for vulnerabilities is allowed, but requests to generate exploits or conduct penetration attacks trigger fallback.
Unlike Fable 5, Claude Opus 5 does not fall back on biology, chemistry, and life sciences questions, as Anthropic considers the model less dangerous in these areas.
Andrew Ng noted that open-weight models have gained broad support on social media but have not yet won in Washington or state legislatures, so continued efforts are needed.
Security officials are frustrated that closed models refuse to assist in security reviews, as attackers using AI agents can discover vulnerabilities faster, further highlighting the importance of open-source models and open-source harnesses.
This article is primarily based on The Batch (Andrew Ng) news summary, which includes first-hand experiential descriptions and third-party benchmark results. Benchmark data comes from Artificial Analysis, but no original report link is provided, so some data should be considered as source claims.
The refusal of closed models in security reviews may prompt more organizations to turn to open-source models, while the release of Claude Opus 5 demonstrates that the trade-off between performance and safety continues.
Primary report
Primary source