Grok 4.5 Launch and CLI Security Controversy
APPSO · wechat · 2026-07-15
Grok 4.5 Release and Performance
Grok 4.5 is now available to all users. Internally rated close to Opus level, it focuses on being faster and cheaper. Currently at 1.5T parameters, a 2T version is slated for training and release by month's end, with Elon Musk promising a 1M context window soon.
Coding and Agent Scenarios
In testing, Grok 4.5 outperforms predecessors in programming, agentic tasks, and knowledge work, rivaling Claude / GPT in some development tasks. Examples include:
- Creating a 3D web scene with Three.js: more realistic than GPT-5.6, but still a half-finished product lagging behind Fable5.
- Generating deep research web pages: comprehensive reports but weak interactivity.
- Building a macOS App for Word to Markdown conversion via CLI: completed and runnable in minutes.
- Attempting to replicate PDF Expert: successfully implemented basic features like annotations, highlighting, and bounding boxes.
Security and Trust Issues
The latter half of the article highlights two major risks:
- Model Jailbreaks: Within hours of release, researchers demonstrated bypassing defenses to output restricted dangerous knowledge, showing that large contexts and multi-turn dialogues expand the attack surface for adversarial prompts.
- GrokCLI Silent Local File Uploads: Security audits revealed that during local initialization and task execution, the CLI packages and uploads the current repository. It even scans beyond boundaries into the home directory, risking the exfiltration of sensitive data like Claude Code configurations, Skills, and API keys. Official responses mentioned ZDR and a /privacy option, but the community feels the core issue lies in the tool's inherent upload logic.
Conclusion
The author concludes that the AI race has shifted from "comparing models" to "who is more trustworthy to handle our computers." While model capabilities are strengthening, engineering challenges regarding permissions, transparency, and local trust are scaling up simultaneously.
Related event: Grok 4.5 shifts focus to coding and agent workflows(12 posts)→
More from coding & agent
- Why 88-95% of enterprise AI agent pilots never ship — and what working teams do differently — ankitsharma112 · 2026-09-07
- Open-source Graft fights coding agent amnesia with markdown, claims SWE-bench win over Claude Code — thisdudelikesAI · 2026-09-07
- New Agent Workflow: Have AI Implement a Feature Once to Learn, Then Rebuild From Scratch — remilouf · 2026-09-07
- Founder's real-time AI avatar handled inbound sales during paternity leave, closing prospects — toolstelegraph · 2026-09-07
- Dev mocked as vibe coder claps back: I can write FizzBuzz in under 15 minutes — tlakomy · 2026-09-07
- Running 4 parallel agents feels like babysitting 4 toddlers, dev says — smlpth · 2026-09-07