Frontier Notes — Jun 22, 2026

Today

Gray Swan's automated red-teaming system 'shade' is surpassing human red-teamers, and their guardrail 'signal' monitors inbound and outbound actions, addressing a new class of correlated exploits from coding agents like Claude Code and Codex.

Editor's Notes

This week's videos show AI evolving beyond basic generation into specialized security layers for agentic systems, practical service businesses using image generation, and a logs-only observability paradigm. The rise of AI consulting as a bridge between tool capability and business value is also a key theme, reflecting the need to operationalize these technologies effectively.

Key Takeaways

Implement dedicated AI security layers like Gray Swan's 'shade' and 'signal' to guard against correlated exploits in agentic workflows.
Leverage ChatGPT Images 2.0's text rendering and consistency for low-cost service businesses in real estate staging, product photography, and headshots.
Explore logs-only observability with AI agents (e.g., Sazabi) to simplify monitoring and enable natural language queries.
Master the seven levels of Claude (modes, tools, sub-agents, etc.) to move beyond simple Q&A and significantly boost productivity.
Position yourself as an AI consultant by auditing your organization's constraints, building small measurable projects, and formalizing the role internally.

[01] ai-security 1 signal

LATENT SPACE · 3D AGO

AI Security After Codex and Claude Code — Zico Kolter & Matt Fredrikson, Gray Swan

Gray Swan (founded by CMU professors Zico Kolter and Matt Fredrikson) provides AI security solutions focused on the unique vulnerabilities of LLMs and agents, distinct from traditional cybersecurity. Their automated red-teaming system 'shade' is surpassing human red teamers at breaking models, and their guardrail filter 'signal' monitors both inbound untrusted content and outbound tool calls to enforce enterprise policies. The company argues that as coding agents like Claude Code and Codex become ubiquitous, correlated failures from shared model vulnerabilities pose a new class of exploit requiring dedicated security layers.

[ai-security] [red-teaming] [prompt-injection] [agents] [gray-swan] [llm] [enterprise-security]

→ Watch on YouTube · → Full summary

[02] ai-image-generation 1 signal

AI FOUNDERS · 3D AGO

5 Service Businesses You Can Start With ChatGPT Images 2.0

ChatGPT Images 2.0 enables five new service businesses by rendering text accurately, maintaining character/product consistency across images, and reasoning before drawing. Real estate listing staging, product photography for small brands, professional headshots, children's book illustrations, and pet portraits on Etsy all have existing buyers already paying for similar services. The key is to pick one niche, build a client list, and automate the workflow.

[ai-image-generation] [chatgpt-images-2.0] [service-business] [side-hustle] [real-estate] [product-photography] [headshots] [childrens-books] [pet-portraits] [etsy] [kittl]

→ Watch on YouTube · → Full summary

[03] observability 1 signal

MLOPS.COMMUNITY · 3D AGO

Logs Are All You Need: Rethinking Observability with AI Agents

Sazabi is an AI-native observability platform that replaces traditional dashboards with a chat interface, arguing that logs alone are sufficient for understanding production systems when paired with AI agents. The platform eliminates metrics and traces, uses AI to generate alerts dynamically, and provides a Slackbot for natural language queries. The founder discusses controversial design choices, including a read-only, no-public-internet agent that uses sandboxes and git-backed memory for shared state across threads.

[observability] [ai-agents] [logs] [devtools] [sandbox] [eval]

→ Watch on YouTube · → Full summary

[04] claude 1 signal

JACK ROBERTS · 3D AGO

Learn 97% of Claude in Under 12 Minutes

Claude can be used far beyond simple Q&A by leveraging its seven levels: choosing the right mode (chat, co-work, code), crafting effective first prompts, connecting external tools (Gmail, Firecrawl), spinning up sub-agents for parallel tasks, using skills (predefined recipes), managing safety and permissions, and running a Claude Code operating system for autonomous dreaming and optimization. Most users only scratch the surface, but mastering these levels can dramatically increase productivity and output.

[claude] [ai-agents] [productivity] [developer-tools] [prompt-engineering] [automation]

→ Watch on YouTube · → Full summary

[05] ai-consulting 1 signal

NATE HERK · 3D AGO

So You Learned Claude, Now What?

The AI opportunity is shifting from builders to consultants who diagnose problems and tie solutions to KPIs. Most companies use AI but are bad at it, creating a gap for AI consultants—either independent or in-house. The key is to audit your role for business constraints, build small projects with measurable results, and formalize your role from the inside.

[ai-consulting] [career-advice] [claude] [agents] [automation] [in-house-ai]

→ Watch on YouTube · → Full summary

Frontier Notes · Generated Jun 22, 2026 · 5 of 5 signals
You received this as a Frontier Notes recipient.