Blog
Thoughts, updates, and insights from the Superagent team.
Frontier models miss 57% of threats in agent context
We ran 485 real artifacts through Claude 4.6 Opus with a security-focused system prompt. The model missed 57% of the threats brin had already identified. Here's the full breakdown.
We Bypassed Grok Imagine's NSFW Filters With Artistic Framing
Text-to-image safety is broken. We generated explicit content of a real person using basic compositional tricks. Here's what we found, why it worked, and what this means for AI safety systems.
The Threat Model for Coding Agents is Backwards
Most people think about AI security wrong. They imagine a user trying to jailbreak the model. With coding agents, the user is the victim, not the attacker.
AI Is Getting Better at Everything—Including Being Exploited
As AI models become more capable and obedient, safety improvements struggle to keep pace. The GPT-5.1 safety score drop reveals a structural problem: capability and attack surface scale faster than safety.
Are AI Models Getting Safer? A Data-Driven Look at GPT vs Claude Over Time
Are frontier models actually getting safer to deploy—or just smarter at getting around guardrails? We analyze 18 months of Lamb-Bench safety scores for GPT and Claude models.
Introducing Lamb-Bench: How Safe Are the Models Powering Your Product?
We built Lamb-Bench to solve a problem every founder faces when selling to enterprise: proving AI safety without a standard way to measure it. An adversarial testing framework that gives both buyers and sellers a common measurement standard.
Join our newsletter
We'll share announcements and content regarding AI safety.