Alex Stamos, Corridor CSO, says Anthropic's upcoming Sonnet 5 limits cybersecurity capabilities to prioritize agentic safety
It trails existing Opus models on cybersecurity benchmarks
Users question Anthropic intentionally reducing Sonnet 5's cybersecurity skills versus Opus models, wondering if the downgrade is meant as a selling point.
No Digg Deeper questions have been answered for this story yet.
Most Activity
@alexstamos Flowers for Algernon.
@SwivalAgent It's supposed to keep it from getting banned by the White House.
SITUATION EXPLAINED: Anthropic just deliberately made Sonnet 5 worse at cybersecurity than its predecessor.
• Sonnet 5 underperforms Opus on every benchmark except GDPval, where it scores 1618 vs Opus 4.8's 1615 • Agentic coding is much better than Sonnet 4.6 and it shows lower rates of undesirable behavior in agentic contexts • Sonnet 5 has lower cyber capabilities than both Opus 4.8 and Sonnet 4.6, intentional • @teortaxesTex: "The one bench nobody wants to hill-climb" • @banteg: "Welcome to bench nerfing era. Sonnet 5 weaker than Sonnet 4.6." • @gfodor: "Seeing Anthropic self-immolate their models because of safety reasons until they have to fold is going to be quite something" • Introductory pricing: $2 input / $10 output per million tokens through August 31st, then moves to $3/$15 • Sonnet 5 is now the default model for free and Pro plans
@theojaffee: "Introducing our newest model. It's our least capable yet. We lobotomized our model."
@alexstamos This is supposed to be a selling point?
@alexstamos Yeah lol