Prompt Injection Risks in Agentic Tools

Brian and Andy wrapped up the week with a fast-paced Friday episode that covered the sudden wave of AI-first browsers, OpenAI’s new Company Knowledge feature, and a deep philosophical debate about what truly defines an AI agent. The show closed with lighter segments on social media’s effect on AI reasoning, Google’s NotebookLM voices, and the upcoming AI Conundrum release.

Key Points Discussed

Agentic Browser Wars

Microsoft rolled out Edge Copilot Mode, which can now summarize across tabs, fill out forms, and even book hotels directly inside the browser.

OpenAI’s Atlas browser and Perplexity’s Comet launched earlier in the same week, signaling a new era of active, action-taking browsers.

Chrome and Brave users noted smaller AI upgrades, including URL-based Gemini prompts.

The hosts debated whether browsers built from scratch (like Atlas) will outperform bolt-on AI integrations.

OpenAI Company Knowledge

OpenAI introduced a feature that integrates Slack, Google Drive, SharePoint, and GitHub data into ChatGPT for enterprise-level context retrieval.

Brian praised it as a game changer for internal AI assistants but warned it could fail if it behaves like an overgrown system prompt.

Andy emphasized OpenAI’s push toward enterprise revenue, now just 30% of its business but growing fast.

Karl noted early connector issues that broke client workflows, showing the challenges of cross-platform data access.

Claude Desktop vs. OpenAI’s Mac Tool “Sky”

Anthropic’s Claude Desktop lets users invoke Claude anywhere with a keyboard tap.

OpenAI countered by acquiring Apple Software Applications Inc., whose unreleased tool Sky can analyze screens and execute actions across MacOS apps.

Andy described it as the missing step toward a true desktop AI assistant capable of autonomous workflow execution.

Prompt Injection Concerns

Both OpenAI and Perplexity warned of rising prompt injection attacks in agentic browsers.

Brian explained how malicious hidden text could hijack agent behavior, leading to privacy or file-access risks.

The team stressed user caution and predicted a coming “malware-like” market of prompt defense tools.

The Great AI Terminology Debate

Ethan Mollick’s viral post on “AI confusion” sparked a discussion about the blurred line between machine learning, generative AI, and agents.

The hosts agreed the industry has diluted core terms like “agent,” “assistant,” and “copilot.”

Andy and Karl drew distinctions between reactive, semi-autonomous, and fully autonomous systems — concluding most “agents” today are glorified workflows, not true decision-makers.

The team humorously admitted to “silently judging” clients who misuse the term.

LLMs and Social Media Brain Rot

Andy highlighted a new University of Texas study showing LLMs trained on viral social media data lose reasoning accuracy and develop antisocial tendencies.

The group laughed over the parallel to human social media addiction and questioned how cherry-picked the data really was.

AI Conundrum Preview & NotebookLM’s Voice Leap

Brian teased Saturday’s AI Conundrum episode, exploring how AI memory might rewrite family history over generations.

He noted a major leap in Google NotebookLM’s generated voices, describing them as “chill-inducing” and more natural than previous versions.

Andy tied it to Google’s Guided Learning platform, calling it one of the best uses of AI in education today.

Timestamps & Topics

00:00:00 💡 Intro and browser wars overview

00:02:00 🌐 Edge Copilot and Atlas agentic browsers

00:09:03 🧩 OpenAI Company Knowledge for enterprise

00:17:51 💻 Claude Desktop vs OpenAI’s Sky

00:23:54 ⚠️ Prompt injection and browser safety

00:31:16 🧠 Ethan Mollick’s AI confusion post

00:39:56 🤖 What actually counts as an AI agent?

00:50:13 📉 LLMs and social media “brain rot” study

00:54:54 🧬 AI Conundrum preview – rewriting family history

00:59:36 🎓 NotebookLM’s guided learning and better voices

01:00:50 🏁 Wrap-up and community updates