OpenAI and Anthropic’s product leads trade subtle jabs: Your models have only just caught up to ours.

Beating AI News: Claude Code head Boris Cherny posted that GPT-6 Astra’s prompt injection protection is “roughly on par” with Gemini Flash and Claude Opus 4.8, adding that his own Claude models had “practically solved” the problem around two months ago. Prompt injection refers to hiding malicious instructions in web pages, documents, and other content to induce AI agents to leak data or perform dangerous actions. Boris also stated that publicly evaluating and naming other labs is an effective way to push them to develop safer models: “We will continue doing this until other labs prioritize safety more.” The post was quickly countered with a Community Note on X. Independent testing by Gray Swan found none of the evaluated models are fully immune to prompt injection, with Claude also having successful attack cases. Another security researcher used a custom web page to target Claude Code Opus 5 Auto Mode, achieving a 60%–80% success rate in small-scale tests. OpenAI’s core product lead Thibault Sottiaux then directly parodied Boris’s phrasing. He said he was pleased to see Claude Code’s new background computer operations finally catch up to Codex: “And it’s our version from May this year.” Boris later acknowledged the original post was “more sarcastic than he intended,” stressing he was genuinely recognizing Astra’s progress. He also clarified that the claim that prompt injection was “solved” refers to the combined performance of Claude models, injection detection mechanisms, and Auto Mode—not that the model itself is inherently immune.