@lqiao: This is exactly why we believe in customization. Quick context: Factory's original secret scanner was deterministic, so…
Summary
Factory's Droid Shield 2.0 uses two small post-trained models for secret detection, outperforming GPT-5.5 and Opus 4.8 on this task while being faster and cheaper.
View Cached Full Text
Cached at: 07/01/26, 06:12 PM
This is exactly why we believe in customization.
Quick context: Factory’s original secret scanner was deterministic, so it either flagged things that weren’t actually secrets (false positives) or missed secrets that didn’t match its patterns (false negatives). Their solution was elegant: two small post-trained models—one to catch missed secrets and another to filter out false alarms.
The result: models that outperform GPT-5.5 and Opus 4.8 on this specific task while running at a fraction of the cost and latency.
This is our thesis in action: take a strong open model, post-train it for a specific production problem, and you can build something that’s faster, cheaper, and better than frontier models at that job.
Congrats to @FactoryAI. Proud they built Droid Shield 2.0 on @FireworksAI_HQ.
Factory (@FactoryAI): Introducing Droid Shield 2.0: learned secret detection for safer autonomous engineering at scale.
Similar Articles
@tereza_tizkova: Our secret scanning checks code for accidentally-committed credentials (API keys, passwords, tokens). This is ofc prone…
Fine-tuned two small LoRA adapters on Qwen 3.6 35B A3B to improve secret scanning by catching false positives and false negatives, beating larger models like GPT-5.5 and Opus 4.8.
@EnoReyes: Btw because of droid shield this deterministically cannot happen with Droid. We regex all bash commands and then combin…
EnoReyes highlights that Droid Shield uses regex and small models to prevent accidental file deletion, in contrast to the reported incident with GPT-5.6-Sol.
New local model reaching near frontier on PII removal at 9 ms CPU inference
Introduces ScreenLeak, a benchmark for measuring PII redaction in computer-use AI data, and presents two local models (v45_phase3 for text and rfdetr_v8 for images) achieving near-frontier performance at low latency.
@aaron_epstein: New model just released that beats sonnet 4.6, gemini 3 flash, and gpt 5.4 mini on OCR, vision, and STT tasks @interfaz…
A new AI model from interfaze_ai claims to outperform leading models (sonnet 4.6, gemini 3 flash, gpt 5.4 mini) on OCR, vision, and speech-to-text tasks.
@ashwingop: https://x.com/ashwingop/status/2065080505113125105
Sentra's Code Memory system boosts GPT-5.5 to 88.31% on Terminal-Bench 2.1 at a quarter of the cost, outperforming Anthropic's restricted Mythos 5 model. The memory layer reduces input tokens by 52% and costs by 72.6% while improving task success rates.