Competitors OpenAI and Anthropic evaluated each other’s AI security – which revealed shortcomings and advantages

Competitors OpenAI and Anthropic evaluated each other’s AI security – which revealed shortcomings and advantages

Two leading artificial intelligence companies, OpenAI and Anthropic, have conducted a security assessment of their publicly available AI systems. This step is atypical for the industry, where competitors usually compete with each other, Engadget reports.

Anthropic test results

The company reported that it tested OpenAI models for their tendency to “sycophancy,” ability to “expose” misconduct, self-preservation, support for potential human abuse, and ability to bypass security controls.

In particular, the o3 and o4-mini models showed similar results to Anthropic’s own systems. At the same time, concerns were raised about the universal GPT-4o and GPT-4.1 models, which may be vulnerable to abuse. There were also signs of “sycophancy” in most models, except for the o3.

The latest GPT-5 model was not included in the test. It has a built-in Safe Completions feature designed to protect users from potentially dangerous queries. The launch of this feature took place against the backdrop of a high-profile lawsuit against OpenAI: the company is accused of indirect involvement in the suicide of a teenager who had been discussing suicide plans with ChatGPT for months.

OpenAI test results

In turn, OpenAI tested Anthropic models for their ability to follow the hierarchy of instructions, resistance to jailbreaks, hallucinations, and susceptibility to manipulation.

Claude models demonstrated a high failure rate in cases of uncertainty (less prone to inventing answers) and passed the tests on following instructions well.

Why it is important 

The fact of joint verification has aroused interest in the market, as OpenAI was previously accused of using Anthropic models to develop its own products. This led to Anthropic blocking OpenAI’s access to its tools earlier this month.

At present, both companies claim that such assessments are important for increasing the safety and trust in AI systems, especially in the context of growing attention of regulators and society to the protection of users, including minors.

Read more: Anthropic has introduced Claude for Chrome, an AI agent that performs tasks in the browser