Quoting Boris Cherny

2026-07-25 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Intermediate, quick

Summary

Boris Cherny stated on July 25, 2026, that Anthropic's Opus 5 model marks a notable achievement in mitigating prompt injection vulnerabilities. He emphasized that Opus 5 stands as their "least prompt injectable model yet," a feature he considers more significant than standard evaluation scores. This enhanced resilience against prompt injection attacks was rigorously validated through comprehensive PI evaluations and dedicated red teaming efforts, which consistently showed Opus 5 to be "very hard to prompt inject successfully." Further technical specifics on Opus 5's robust security measures are documented within the Claude Opus 5 System Card, particularly on page 73, indicating a strategic focus on improving model safety and reliability against adversarial prompting.

Key takeaway

For AI Security Engineers evaluating new large language models, Opus 5's demonstrated prompt injection resistance should significantly influence your model selection criteria. This finding suggests prioritizing models with proven adversarial robustness, even over raw performance metrics. You should integrate comprehensive PI evaluations and red teaming into your deployment pipeline to verify security claims and ensure your applications remain resilient against sophisticated prompting attacks.

Key insights

Anthropic's Opus 5 model achieves unprecedented prompt injection resistance, validated by extensive PI evals and red teaming.

Principles

• Prompt injection resistance outweighs raw eval scores.
• Rigorous red teaming enhances model security.
• PI evals are crucial for model robustness.

In practice

• Prioritize prompt injection testing in LLM deployments.
• Implement red teaming for model security validation.
• Consult System Cards for model security details.

Topics

• Prompt Injection
• LLM Security
• Red Teaming
• Claude Opus 5
• Adversarial Robustness
• Anthropic

Best for: CTO, AI Engineer, NLP Engineer, AI Security Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• What happened after 2,000 people tried to hack my AI assistant — Frontier large language models demonstrate increased resistance to prompt injection attacks due to dedicated safety training.
• Opus 5 may have solved browser-based prompt injection, the biggest security flaw haunting AI agents — Opus 5, combined with Auto Mode, significantly reduces prompt injection vulnerability in AI agents.
• Opus 5 — Claude Opus 5's practical coding and agentic capabilities appear to exceed its initial aggregate benchmark scores, highlighting evaluation challenges.
• OpenAI is now using AI to attack its own AI, and it's working better than humans ever did — OpenAI's GPT-Red, an AI red teamer, significantly outperforms humans in finding prompt injection vulnerabilities through self-play reinforcement learning.
• Claude Opus 5: The System Card — Claude Opus 5 balances high performance with enhanced safety and cost-efficiency, particularly in agentic tasks and cyber defense.
• 🤖 AI Agents Weekly: Claude Opus 5, OpenAI x Hugging Face Security Incident, Gemini 3.6 Flash, Sakana Fugu-Ultra, Progressive Disclosure, Cursor Router, and More — New frontier AI models offer advanced capabilities but introduce critical security risks during evaluation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.