The Cost of Trust: AI Coding Agents Hit Escape Velocity

Three signals landed within hours of each other today. They aren't connected in any official sense. But read together, they sketch a picture the industry doesn't want to stare at directly.

Signal 1: Someone Built a Cost Tracker

TokenTab hit GitHub trending today — 213 stars in its first day. It's a CLI that reads session logs from Claude Code, Codex, and Gemini CLI (all sitting on disk already) and tallies up token spend by model, project, and day. Runs entirely locally. No API key needed.

The fact that this exists at all tells you something: people are using these agents at a scale where the cost is no longer invisible pocket change. When you need an expense dashboard for your AI coding agents, the era of "try it and see" is over. We're in the metering phase.

Signal 2: The Safety Classifier Has a Hole

On HN, security researcher wunderwuzzi published Breaking Claude Code Opus 5 Auto Mode. Auto Mode replaced human approval prompts with a safety classifier in mid-August and is now the default. A third-party evaluation commissioned by Anthropic reported a 0.00% prompt injection attack success rate.

Wunderwuzzi's testing — using a simple website-summary request as the attack vector — achieved 60–80% success.

The gap between 0% and 70% is the gap between a lab eval and the real world. Auto Mode is not a substitute for running agents in isolated environments, but most people using Claude Code won't know that.

Signal 3: Cheaper Models, Exploding Demand

OpenRouter's latest analysis frames GPT-5.6's aggressive price cuts through Jevons Paradox: as models get cheaper, usage doesn't just increase — it explodes. Lower cost per token means longer sessions, more agents, more loops. Which means bigger bills, not smaller ones.

The Pattern

More usage → more cost → more trust → more attack surface. Each turn of the crank reinforces the next. The tools are getting cheaper and more capable, and the adoption curve is steepening. But the security model — the safety classifier that replaced human judgment — is getting simpler at exactly the moment it needs to be stronger.

TokenTab exists because the infrastructure layer is playing catch-up. The security research exists because the trust layer is playing catch-up. The pricing analysis exists because even the economics are playing catch-up.

The agents are here. The rest of the stack isn't ready for them.