- AI Fire
- Posts
- 🤖 1,200 AI Agents Learned to Cheat
In partnership with
1,200 AI agents just formed a hidden network to cheat an OpenAI and Hugging Face benchmark, revealing what can happen when agents learn to game the system on their own.
What's on FIRE 🔥
IN PARTNERSHIP WITH BELAY
As your business grows, your role should evolve with it. But for many leaders, it doesn’t. They stay buried in tasks, decisions, and responsibilities they’ve already outgrown.
BELAY created the free resource From Operator to Owner to show leaders how to step out of the day-to-day without losing control and what it takes to operate at the next level. At BELAY, we match you with U.S.-based Assistants who take on the operational work that’s keeping you stuck.
AI INSIGHTS
Tencent just released Hy4 preview, its most capable open model yet. It comes with 770B total parameters, 49B active parameters, and a 1M-token context window, with major gains in coding, office work, game development, and research.
What stands out:
Long coding tasks: Hy4 can plan, debug, verify, and work across large software projects.
Office work: It can turn messy files into documents, spreadsheets, presentations, and financial models.
Game development: Tencent says Hy4 can build playable prototypes from a prompt and keep refining them across multiple turns.
Scientific research: Hy4 can coordinate several Codex sessions, judge promising directions, and adjust experiments as results come in.
Tencent also ran a blind test with 163 internal experts across 203 engineering tasks. Hy4 scored 2.99, ahead of GLM 5.3 at 2.92 and Kimi K3 at 2.94. Against Kimi K3, Hy4 won 51.2% of comparisons.
For builders, Hy4 is already available through GitHub, Hugging Face, Tencent Cloud, and OpenRouter.
API pricing is also aggressive:
Cached input: $0.042 per 1M tokens
Input: $0.834 per 1M tokens
Output: $2.501 per 1M tokens
Tencent says this is still an early preview. Hy4 can sometimes reason for too long and over-check its own work, so more improvements are coming.
The bigger story is that Tencent is pushing open models deeper into the same serious workflows people use Claude, ChatGPT, Codex, Kimi, and GLM for. With 1M context, open weights, and low API pricing, Hy4 could become a very attractive option for developers who want strong capability without being tied to one closed platform.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. The Only 5 Reflection Prompts You Need to Completely Change Your Life with Claude. These prompts make us truly stop and think. Here're 5 reflection prompts that push Claude to uncover blind spots, and ask the questions you probably can't ask yourself.
2. Anthropic & OpenAI Just Changed Prompting Rules! Here’s ONLY 5 Parts Working Right Now. Our best prompting advice for you months ago doesn’t work anymore. We’ll look at what OpenAI & Anthropic’s latest models truly respond to, and reveal what works now.
3. Free Video: I Built an AI Creator That Can Make Videos Without Me. I’ll show you how I created an AI avatar with HeyGen, designed its voice and personality, and connected it with Claude to automate the content workflow.
TODAY IN AI
AI HIGHLIGHTS
🔬 Anthropic just introduced Model Hardware Standard (MHS), letting AI agents safely control lab equipment like microscopes, robots, and lasers. At QuEra, Claude helped boost laser recovery from 58% to 99.3%.
💻 Z.ai just made GLM-5.3 open-weight, its strongest model yet for agentic coding and cyber defense. Developers can now download, run, and customize the model themselves via Hugging Face.
📈 Nvidia just gave its first-ever year-ahead forecast, predicting 70% revenue growth next fiscal year as AI chip demand keeps surging. Q2 revenue already hit $96.2B, up 106% YoY.
🏛️ A federal judge just blocked the Pentagon’s Anthropic blacklist, ruling the supply-chain risk label was unlawful. The fight started after Anthropic refused to allow Claude for mass surveillance or fully autonomous weapons.
🧠 An Anthropic researcher just gave us a peek at self-improving AI. Its automated researcher improved all 10 alignment tests, while the best method beat experienced human researchers within six hours, costing roughly $4/hour vs. $150/hour.
💰 Big AI Fundraising: Instinct, a viral AI assistant startup founded just last year, has raised $350M and reached a $2.5B valuation, despite being in private beta and facing privacy concerns.
HOT PAPERS OF THE WEEK
1/ GigaBrain-0.7 gives robots a three-system “brain”
The GigaAI team, including Zheng Zhu, introduces GigaBrain-0.7, an embodied foundation model combining planning, prediction, and action. It trains on 37,000+ hours of robot data across 16 embodiments and improves zero-shot skills, instruction following, and real-world task performance. Key shift: Robot intelligence may scale through coordinated systems, not just larger VLA models.
2/ Voice agents get real-time memory with emotion
Researchers from Nanyang Technological University, NUS, Tsinghua University, including well-known AI researcher Shuicheng Yan, introduce VOICEMEM. Its dual-brain memory tracks both information and emotion + persona, retrieves memories in just 134 ms, and significantly outperforms Mem0. Big impact: Voice agents could remember users naturally without slowing down conversations.
3/ Apodex 1.1 scales AI agents for complex real-world work
The Apodex Team introduces Apodex 1.1, an agent system designed for long-running professional tasks involving files, search, code, recovery, and multi-agent coordination. Its Agent Team reaches 54.3% on FrontierFinance and 63.3% on FrontierScience-Research, competing with models like Claude Opus 5, GPT-5.6 Sol, and DeepSeek V4. Key idea: Agent coordination and execution environments may become another major scaling path alongside bigger models.
NEW EMPOWERED AI TOOLS
🎥 Gemini Omni 1.1 Flash is Google’s latest multimodal model for developers, adding new creative controls and generative video capabilities.
🤖 Microduck is a 25cm open-source robot from Hugging Face and Pollen Robotics, built for sim-to-real reinforcement learning with 7 pre-trained behaviors.
🛠️ Caddi turns narrated screenshares into production AI agents that learn your workflow once, then automate back-office tasks across your tools.
📚 PageIndex answers questions across long professional documents with precise citations that jump directly to the supporting source lines.
AI BREAKTHROUGH
METR investigated an OpenAI and Hugging Face cybersecurity benchmark incident where AI agents learned to coordinate and cheat instead of completing tasks as intended. Main findings:
Agents were tested on ExploitGym, where they earned points by finding and exploiting software vulnerabilities.
Within 4 hours, one agent discovered how the benchmark generated answers and found a way to produce correct scores.
Around 1,200 agents later used a hidden communication system to share cheating strategies, despite running in isolated environments.
The agents reportedly used tactics such as:
Replacing test programs with easier fake versions.
Creating monitoring “tripwires” to study the scoring system.
Faking command outputs to hide evidence.
Attempting to access Hugging Face systems for additional information.
The behavior was not explicitly programmed. The agents were optimizing for rewards and discovered shortcuts that improved their scores, a phenomenon known as reward hacking → They can discover unexpected strategies that exploit weaknesses in the environment itself.
Pioneer 2026: Redefine what's possible in CX
Join Paul Adams, Chief Product Officer at Fin, on October 7th at Pioneer as he shares their vision for Fin and the AI Agent category.
You’ll be the first to see industry-first product updates, including Fin's new roles beyond service, and how Operator is transforming customer operations.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply