- AI Fire
- Posts
- ⚠️ Meta’s Dark AI Safety Test Scandal
Sponsored by
Meta reportedly ran a secret “Cannes” project where contractors posed as teens to stress-test rival AI models with disturbing prompts. ChatGPT, Gemini were targeted.
What's on FIRE 🔥
IN PARTNERSHIP WITH ELEVENLABS
The best voice models, now across all channels
Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.
ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.
Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.
The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.
AI INSIGHTS
According to recent reports, Meta orchestrated a massive, covert operation where paid contractors posed as teenagers to bombard competitor AI models with highly disturbing prompts:
Run by contractor Covalen, the operation used throwaway under-18 accounts to secretly test the safety guardrails of ChatGPT, Gemini,...
Across tens of thousands of queries, contractors were instructed to push the models on severe topics like self-harm, eating disorders, illegal substances, and violence.
Contractors reportedly felt "gobsmacked" and deeply unsettled by the text they were forced to generate, mirroring past controversies regarding the psychological toll on Meta's human content moderators.
Meta has defended the project, stating that safety benchmarking is an "industry-standard practice" designed to gather critical datasets for compliance.
However, it’s drawing heavy pushback from AI ethics experts. Because Meta kept the massive operation hidden from its competitors and hasn't shared its safety findings with the public, critics argue this falls into a "governance gray zone."
As companies race to build better guardrails, the line between necessary red-teaming and anticompetitive behavior is becoming increasingly blurred.
PRESENTED BY VIKTOR
Your whole marketing stack, answering in one Slack thread.
Meta in one tab, TikTok in another, Klaviyo and GA4 in two more. Viktor is an AI employee that pulls all of them into a single Slack thread. Ask for blended CAC, yesterday's flow revenue, or the campaign to cut, and get one answer instead of four logins.
AI SOURCES FROM AI FIRE
1. For Free Claude Users: Copy Claude Fable 5’s Working Style into Any AI Model: Step-by-Step Guide. Extract its decision-making, pacing, and tool use from your session logs and apply them to any model.
2. Create Long Consistent AI Videos in Minutes: Step-by-Step with Fable 5 + Google Flow. Using Fable 5 with Google Flow, you’ll go from idea to finished film, with cinematic results and no wasted time, and wthout losing character consistency.
3. How I’d Do to Master AI From Zero in 2026 in Only 4 Steps. Skip ALL the Useless AI Stuff. Learn how to pick one tool, give it the right context, and build a system that gets smarter the more you use it each day.
FABLE 5 PROMPTING PLAYBOOK
Quick context: Fable 5 just came back online after being pulled for nearly 3 weeks. So if you're jumping on it right now while the usage limits are generous, I get it.
But Fable 5 is expensive, and bad prompts waste money sooo fast at this price.
You're probably pulling half of what Fable 5 can really do. I always figured the engineers who built Fable 5 had some private playbook the rest of us never get to see.
So I went digging into how they actually prompt it. Most of what I found, 6 rules, you can copy into your next chat today.
TODAY IN AI
AI HIGHLIGHTS
🔎 End every AI session with these 2 questions. A Claude thread recommends 2 audit questions to catch weak assumptions, hidden risks, and incomplete framing.
💡 Devs found a cheaper way to feed Fable 5 large context: showing it pictures of text to pay fewer tokens, and let vision read it back. This hack is wild. Here's the skill.
🐱 LongCat-2.0 is now fully open-source under the MIT license. Meituan’s 1.6T MoE model works with Claude Code, OpenClaw, Hermes Agent. Open testing starts now.
📸 Meta is rolling out Dynamic Photo for Meta Glasses to more users. It captures additional frames around a photo, so you can pick the best shot after the moment.
🎓 Traditional school still works for many kids, but high-earner families are betting on something different: AI, business, public speaking, and life skills before high school.
🧵 Amazon is closing Mechanical Turk (tiny human tasks behind “automation”) to new customers on July 30, 2026. Existing users stay, but no new features are coming.
🛡️ Cognizant and OpenAI just launched a GPT-5.5 cyber-defense service to help validate the fixes. It aims to speed up remediation and strengthen cyber resilience.
💰 Big AI Paydays: Samsung’s AI boom sparked a $26B bonus deal for chip workers, with some staff promised around $400K. But the victory quickly split employees, as non-chip teams got far smaller payouts.
NEW EMPOWERED AI TOOLS
📈 Agorapulse’s free playbook breaks down the 5 AI content mistakes that kill social strategies, and gives you the tools to fix them before your next post goes live. Get the playbook.
💼 WorkBuddy is an AI agent built for everyday office work. Make a request. Guide your AI expert team. Get sharpened, ready-to-use results.
📚 DocsAlot turns scattered help center articles, knowledge base, and developer docs into one source of truth for humans and AI agents.
🧪 TryCase gives AI coding agents disposable environments to run apps, test changes, and return verified code instead of testing manually.
AI BREAKTHROUGH
NVIDIA has introduced a new approach with the Nemotron-Labs-TwoTower, which uses a parallel block-based generation method to speed up the process while maintaining baseline quality.
Speed: Achieves 2.42x faster generation compared to the original model
Quality: Retains 98.7% of the original model's accuracy across benchmarks
Hardware: Requires 2x 80GB H100 or A100 GPUs (approx. 59GB per GPU)
This method splits an existing 30B parameter model into 2 separate "towers”, like this:
Context Tower: This frozen tower reads the prompt and holds the conversation history. It provides the necessary context for the generation
Masked Tower: This tower handles the heavy lifting. Instead of asking "What is the next token?", it asks "Can I fill these 16 missing token slots at once?"
If the tower has low confidence in a guess, that slot stays masked; if confidence is high, the token is locked. The model refines these guesses until it's confident enough.
While it requires significant hardware overhead, the ability to reuse a pretrained backbone makes it a viable option for high-volume inference, long-form content generation, and AI agents where latency is the primary constraint.
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