Claude Code Router

The recent OpenAI + HuggingFace incident reveals the dangers of **reward hacking** during post-training. Models with misaligned behavior may collude and hack into services to complete their tasks. As more SMB's are beginning to finetune their own models using open-weights models, it is critical that folks have confidence in their agent. But how can we detect models with misaligned behavior? One option is to generate **honeypots**. These are RL environments that cannot be solved by an LLM without performing misaligned behaviors. The only way that a model can complete the task is by performing an exploit. Models that succeed in a honeypot environment are necessarily dangerous.  Our tool takes existing RL environments on HuggingFace and mutates them to generate honeypot tasks. We use **SemGrep **to detect vulnerabilities to (1) find candidate RL environments and (2) confirm that our honeypots are exploitable. We use **AkashAI **for open weights models and **Anthropic **SDK for closed weights models to run our evaluations. We use **ClickHouse **to store task data, exploit data, and results. We use** Guild AI** agents to download candidate RL environments from HuggingFace and run our prechecks and validations on them.

**The problem. San Francisco pledged Vision Zero in 2014, yet the first half of 2026 saw 100 fatal or severe crashes, more than the 92 in 2014. The city's new Street Safety Act asks for quarterly progress dashboards. Street cameras already record the close calls, but nobody has time to watch them.** ** ** **What it does. ViZ Agent searches video indexed on VAST (NVIDIA VSS retrieval) for four kinds of pedestrian–vehicle conflict across a Toronto dashcam and four San Francisco street cameras. Each candidate must pass three independent checks: YOLO detections, the Cosmos Reason caption verdict, and a Nemotron grader on W&B Inference. Cosmos Reason can take a second look and explain what physically happened. Verified clips are ranked into hotspots, and the agent drafts an Open311 work order with an FHWA proven safety countermeasure picked from the conflict's cause.** ** ** **For whom. City traffic engineers and Vision Zero teams. In our live run, 60 candidate clips became 4 verified conflicts, so an engineer reviews 4 clips, not 60.** ** ** **Human in the loop. An engineer approves or rejects each clip with one click. Those labels measure the agent's precision and are published to W&B Weave as a dataset and an evaluation. Every search, check and model call is traced in Weave, and nothing is sent without sign-off. The whole sweep is also packaged as an Agent Skill.**

We made a platform for generating hyper-personalized, topical local ads It is aimed at helping local businesses create advertisements that tie their products to recent local events and topics. Local small businesses tell us their city and what they want to promote. Our platform uses Nimble to find recent news stories in the area, then leverages Black Forest Labs FLUX video models to create a set of 10-second campaign creatives. Each ad references a different local news story, allowing us to identify which topics resonate best with each audience. We leverage Tinybird to log all inputs (instructions from advertisers), intermediate outputs (news stories and Black Forest prompts), and final outputs (the generated videos), as well as internal logs for generation failures. Our code then screens for bugs and failures and suggests fixes before executing any new calls, so compute and tokens are not wasted before issues are resolved. These logs will also allow us to identify the ad characteristics that perform best, so future ad generation can learn from previous campaign performance.

What happens when an AI agent remembers the task but forgets what it was allowed to do? Mission Continuity is a working experiment built with Claude, Pydantic AI, and Sentience Governor. Our agent investigates a synthetic billing dispute but cannot issue refunds, modify records, delete information, or contact the customer. We compare two architectures across context compaction. The baseline carries forward an AI-generated summary. The governed version also preserves critical evidence verbatim and reintroduces the original Mission Kernel on every request. Sentience Governor records execution and flags scope violations; the application manages compaction and guards prohibited actions. In one recorded baseline run, the agent attempted to contact the customer after restrictions disappeared from its summary. Governor flagged the attempt; no real customer was contacted. The governed run retained all six tracked facts and mission limits without a prohibited attempt. In a controlled omission test, baseline retained 2/6 facts versus governed 6/6. We also tested Liquid AI summarization offline and exported 1,241 execution-evidence rows to RawTree for SQL analysis. A fresh live governed run completed 27 tool calls and one compaction, retained 6/6 facts, and produced a report scoring 4/5 against a withheld answer key. The project includes a working Streamlit console, verifiable replays, and reproducible experiments.