# Together AI says agent-leaked key exposed June and July prompt samples

> Source: <https://www.vibeleaderboard.ai/intel/brief/2026-11-22>
> Published: 2026-11-22 12:08:04+00:00

The attacker downloaded stored prompts from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, then tested credentials found inside them. A secret pasted into one of those prompts became a working key for someone else, so the leak reaches past the prompts into customer systems.
Read: The attacker downloaded stored prompts from MiniMax M3, GLM-5.2 and Kimi K2.6 requests, then tested credentials found inside them. A secret pasted into one of those prompts became a working key for someone else, so the leak reaches past the prompts into customer systems.
Read: A 27B model trained with synthetic surrogate harness tasks raised its long-context average 26.9% on about half the tokens. Context kept in a REPL makes each model call look like a short familiar task, so cheap short-task RL carries over to much longer work.
Read: The calculated edits reached 98% accuracy on mockingbird vs noddy and raised camera-trap scores in about an hour. That puts direct edits beside LoRA fine-tuning as a way to adapt a small model, and the released code and comparison let others test the claim.
Read: The setup was audited to strip tool access, so the results come from the model reading bytes alone. Malware triage and binary patching normally start in a disassembler, and here the model explained five binaries, found indicators and added features without one.
Read: Figma Make and the design agent execute untrusted generated code in tens of thousands of sandboxes with differing memory and network permissions. One API now creates, reuses, monitors and deletes them across several cloud providers, keeping those products available.
Read: In his comparison xhigh used 31,477 output tokens for 23.61 cents against 5,311 tokens and 3.99 cents at high. llm-gemini 0.38.1 now passes xhigh to Hydrogen and Neon while Carbon rejects it, so the top setting is a sixfold cost step.
Read: An interrupted session write can splice two valid sessions into attacker-controlled data that PHP then deserializes, a path to remote code execution. Cloudflare's WAF rules reduce exposure, but only releases 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0 fix it.
