Prompt injection disclosures: 4 labs compared

A comparison of Anthropic, OpenAI, Google, and Meta shows there is still no common standard for reporting prompt-injection risk.

Anthropic released the most detailed system card, breaking out four attack surfaces and reporting a 31.5% browser attack success rate before safeguards.

OpenAI gave a single connectors robustness score, Google folded the issue into a broader safety framework, and Meta did not publish a closed-model card.

Security experts say the inconsistent disclosures make it difficult for buyers and security teams to compare models or set a shared benchmark.