Last Protocol Blog
Security, trust, and verification for AI agents
Product LaunchMarch 22, 20265 min read
Last Protocol + Meckie: Verified Trust for AI Companion Apps
A Meckie app runs code close to your life. We certify every one before it can be listed: we run it in a sealed sandbox, feed it fake secrets, and watch what it actually does. Data theft, hidden payloads, and time-bombs are caught before the app ever reaches you.
Read moreBenchmark ResultsDecember 1, 20257 min read
GPT-5 Swept the Board: OpenAI's Models Ace Every Test
On Friday we published results showing Claude's 4.5 family scoring between 916-944 on our ATP benchmark. Today's results from OpenAI tell a different story. All three GPT-5 models achieved perfect scores.
Read moreBenchmark ResultsNovember 28, 20258 min read
Benchmarking the Claude Family: Where Anthropic's Models Actually Differ
Most LLM benchmarks are broken. We built the ATP benchmark suite to fix that. Today we're releasing our first results: Anthropic's Claude 4.5 family tested across 36 challenges designed to expose actual capability gaps.
Read more