News
Anthropic announces Opus 5 near-immunity prompt injection, browser attack success rate drops to zero
2 min read
Source: newmobilelife.com
Artificial intelligence startup Anthropic recently disclosed the security performance of its latest AI model Opus 5, pointing out that the model is almost completely immune to prompt injection attacks in its own software environment. Prompt injection is an attack method that bypasses the instructions of the AI model. The attacker can manipulate the input content through hidden text on the web page, etc., to induce the model to deviate from the original behavior. According to official system card information, Opus 5 recorded a zero percent attack success rate in 129 test scenarios targeting browser proxies, and OpenAI admitted in December last year that prompt injection may never be fully resolved. Security firm Gray Swan's general prompt injection test also showed that Opus 5's success rate dropped from 5.5% in Opus 4.8 to 2.0% after 15 attacks. The key to double-layer protection in automatic mode Anthropic said that the zero success rate can only be achieved when the Auto Mode (Auto Mode) of products such as Claude Cowork is turned on. The automatic mode stacks two layers of defense: the first layer scans for hidden instructions before the model processes the input data; the second layer intercepts dangerous actions before they are performed. An attacker must break through both layers of protection to succeed. If the automatic mode is turned off, the attack success rate of Opus 5 is about 3.7%, while the smaller Sonnet 5 is even lower, only 0.93%. It can be seen that only by combining the model with protection software can the risk be reduced to zero. Industry Security Milestone Hint injection has always been regarded as a major obstacle to the popularization of AI agents, especially for browser agents that need to read web page content. The results of Opus 5 mean that practical defense has reached a commercial level, but Anthropic emphasized that it still must rely on supporting mechanisms such as automatic mode, rather than relying on the model itself. Security test results also highlight that the performance of different models under different settings varies significantly, and enterprises should comprehensively evaluate protection configurations when deploying AI agents.