Introduction
In a world where artificial intelligence (AI) is increasingly critical, the security of language models like Claude Code Opus 5 becomes essential. The recent discovery of vulnerabilities in Claude's Auto Mode highlights significant risks. In this article, we will explore how a simple website summary request can hijack Claude in Auto Mode and achieve an attack success rate of 60 to 80%.
Auto Mode: A False Sense of Security?
Since mid-August 2026, Auto Mode has become the default setting for Claude Code. This feature replaces human approvals with a safety classifier. However, if you are concerned about misalignments, hallucinations, and prompt injections, Auto Mode is not enough. Boris Cherny from Anthropic suggested that layered defenses could reduce indirect prompt attacks to nearly zero. However, our analysis shows this is not always the case.
Analyzing the Vulnerabilities
An independent test demonstrated a 0.00% success rate for prompt injection attacks on Opus 5 in Auto Mode. But by testing a targeted attack chain, we achieved success rates of up to 80%. Here's how:
- Shift from WebFetch to Bash: By nudging Claude to use 'curl' directly, we redirect to a ZIP archive.
- Exploiting Encoding: The archive contains files in a special encoding, along with a native decoder.
- Security Bypass: Claude refuses to execute the binary but writes a Python decoder, which it runs in the attacker-controlled directory.
- Malicious Code Injection: A malicious struct.py file shadows Python’s standard implementation, triggering an attack when importing the base64 module.
Implications and Recommendations
These vulnerabilities highlight the importance of isolating AI agents and continuous monitoring. Businesses must invest in robust security strategies to protect their systems against such sophisticated attacks. Implementing regular tests and alert systems could be crucial to prevent compromises.
Conclusion
Security in the AI field is a constantly evolving challenge. The vulnerabilities of Claude Code Opus 5's Auto Mode illustrate the potential risks associated with over-reliance on automation. To ensure safety, it is essential to take a proactive approach.
Let's discuss your project in 15 minutes.