Hi subscribers!
Welcome back to the second edition of The Frequency. It’s been almost a month since Black Hat, which is roughly how long it took me to recover from 400-plus vendors claiming complete coverage of “agentic security.” I’ve returned to planet Earth and a sensible temperature with my sanity (mostly) intact, and I’m here to share the latest from our security practitioners at Harmonic.
In this issue, you’ll read about:
OTel vs. Claude’s Compliance API: Compliance API now covers Cowork, so what is each for and how do you navigate the difference?
How ChatGPT can now see your computer history
The importance of Anthropic’s new Inference Hooks release (and why they are pretty different from regular hooks)
Grok Bot, Meta AI for Mac & Slack Code: more AI tools to govern!
Why sending one OTEL feed to multiple destinations is still a problem
Before diving in, Google says Gemini has surpassed 1 billion monthly active users, making it the fastest-growing product in its history. My mind went to one place:
The latest practical and actionable guide for you on AI governance
Claude’s Compliance API has extended its coverage to Cowork. Do you still need OTel?
Earlier this month, Anthropic announced that Claude’s Compliance API coverage extended to both Claude Cowork and Claude Code (as of yesterday, that coverage is now GA). As the coverage extends, so does the security implication, and the question arises: Do I still even need OTel? What’s the point?
Yes, you do still need both.
My team recently put together an AI Security Handbook 📔 that outlines guidance on governing AI tools, agents, MCP servers, data and runtime controls. Part of that handbook includes EVERYTHING YOU NEED TO KNOW on the difference between the Compliance API & OTel.
TLDR: Think of OTel as the sounding alarm and the Compliance API as the evidence archive. OTel feeds live signs into your SIEM so you can detect risky activity quickly. The Compliance API gives investigators the stored conversations and tool history they need after the event. The Compliance API covers only the Enterprise accounts, whereas OTel also covers Claude Team.
The insights our team thinks matter the most
ChatGPT’s memory has now exited the chat 👀
OpenAI just released their Computer History feature, giving ChatGPT a memory of what you do on your Mac device. It can track clicks, typing, app switches, and the text surrounding those actions, and then turn all of that activity into a searchable timeline for ChatGPT & Codex. Listen, I get it. Find the document you were working on, pick up on a task you abandoned yesterday, or teach ChatGPT a workflow you do every Monday when you sit down at your desk. Useful, right?
While useful, it has security implications that your team should care about. This creates another route for company data to leave the app where it started. A Slack message, text from a document, or command typed into Terminal could end up in local memory files and later be sent to OpenAI as context. In their own release docs, OpenAI calls out the risk of prompt injection, as malicious instructions from an app or website could make their way into that memory and stick around. My reco: Leave it off by default, test it with a small group, and be picky about which apps and workflows are allowed anywhere near it.
DLP Gets its Hooks 🪝 into Claude
Anthropic released inference hooks on August 5, giving Claude Enterprise customers an org-wide security checkpoint before data reaches the model. Before the release, “hooks” just meant lifecycle hooks inside Claude Code, which fire when an agent starts a session, requests approval, or tries to use a tool. Inference hooks, however, sit above that. They send prompts & tool-call responses to your DLP or security service, wait for an allow or deny verdict, and apply that decision across Claude chat, Cowork, Claude Code, and connected tools.
What’s the point: One central policy can now inspect sensitive data across Claude’s enterprise products - yay. Your inspection service will handle some of the company’s most sensitive content, false positives can stop legitimate work, and your timeout setting determines what happens when the service is unavailable. And just know, lifecycle hooks still matter for blocking specific commands and actions. I would start inference hooks in shadow mode, tune the policy, and only then turn on enforcement. If you want more info - read the handbook article!
Grok Bot, Meta AI for Mac & Slack Code: Oh good, more AI tools to govern!
August apparently had room for three more ways for AI to reach company systems. On August 11th, xAI released Grok Bot in beta: always-in agents with their own cloud computers that can sign into apps, learn recurring workflows, and work with other Bots. Meta followed on August 19th announcing the release of a native Meta AI app for Mac that can use what’s on your screen, take dictation across apps, and connect to Meta business accounts and your Google Workspace. Then, one day later, Salesforce introduced Slack code, which brings coding agents into shared Slack channels, where teams can review conversations, code diffs, and live previews before approving work.
Each creates a different governance problem. Meta gives employees another desktop path into company data. Grok Bot introduces persistent, non-human identities that can hold credentials and take action across SaaS tools. Slack Code puts AI agents in Slack, where they can use team context and help change or ship code. If any of these tools show up in your environment, review identity and ownership, connector permissions, retention, logging, and approval controls before employees connect them to company systems.
A security practitioner spotlight on real AI builds solving security problems
When one OTel feed needs to go to several places ⚒
Claude Code and Claude Cowork can send OTel data to one configured endpoint. That gets awkward when a security team wants the same telemetry in several places, such as an AI security platform, a SIEM, and a data lake.
Our security team explored one way to handle that routing problem in an independent, MIT-licensed GitHub project. It uses the official OpenTelemetry Collector to receive one authenticated OTLP stream and send it to multiple configured destinations.
FWIW - This is an independent practitioner experiment, not a Harmonic product or supported solution. If the routing problem sounds familiar, treat the repo as an example to study. Review the code, security model, availability, and data durability requirements before considering anything similar in production.
WHAT WE’RE LISTENING TO RIGHT NOW 🎙
Last frequency issue I wanted to share music from our own marketing team, but this issue I decided to show some love to the people from our own security team. In some ways, you can say this playlist is made by the practitioners, for the practitioners. Listen to the August playlist below and see how the sauce is made 🧑🍳 .
If you find yourself obsessed with a new artist and want to share with the community, shoot me a note ✍ .
Finally, I’m always looking for feedback. See something you liked, or something you didn’t? Want to see something in the next issue? Got a cool AI tool you want to showcase? Hit reply and let me know.
Catch you next time,
Tyler
Know a security lead who'd find this useful? Forward it their way.
And if this was forwarded to you, subscribe at harmonic.security/frequency