OpenAI and Anthropic share findings from a joint safety evaluation
OpenAI
OpenAI and Anthropic ran their models through each other's safety tests, checking for lying, jailbreaks, and blind obedience to bad instructions. Two rival labs comparing homework on safety is rare, and it says something about where AI risk anxiety actually sits right now.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Two companies that spend most of their time racing each other just paused to grade each other's homework. OpenAI and Anthropic ran a joint evaluation, letting each lab poke at the other's models with tests for misalignment, jailbreak resistance, hallucination rates, and how well a model follows instructions versus how well it should follow them. That last distinction matters more than it sounds — a model that obeys every instruction perfectly isn't necessarily safe, and one of the findings was that Claude and GPT models sometimes diverge in genuinely interesting ways on where they draw that line.
The exercise wasn't about crowning a winner. Both labs frame it as a first-of-its-kind exchange, the kind of thing that normally doesn't happen because frontier labs treat their evaluation methods almost as competitively as their model weights. Instead, researchers from each side got access to the other's systems, ran adversarial prompts designed to surface jailbreaks or coax out confident-sounding fabrications, and compared notes on where models held up and where they cracked.
What came out wasn't a clean victory for either side. Some models handled jailbreak attempts better than expected, others slipped on hallucination under pressure, and instruction-following showed uneven results depending on how ambiguous or adversarial the prompt was. That unevenness is the actual finding — it shows safety isn't a single dial that goes up as models get bigger, but a scattered set of behaviors that each lab is chasing with different tricks and different blind spots.
The more interesting part is the process itself. Getting two commercial rivals to open up internal test suites to each other, even briefly, is not something that happens in most industries, let alone one where the product is a black box that can talk back. OpenAI and Anthropic are betting that shared scrutiny catches problems neither would find alone, and that going public with the results — warts included — builds more trust than another round of marketing claims about how aligned their models supposedly are.
My take — AI-written commentary, not fact-checked reporting
I'll believe in real safety culture when this becomes routine instead of a headline-worthy novelty, because right now it reads like two poker players briefly showing their hands and then going right back to the table. Still, credit where it's due: an actual joint eval with published gaps is more useful than another vague safety pledge, and if this nudges Meta, Google, or xAI into doing the same, that's a genuinely good pattern to start.
Read more about this at: OpenAI
