Your AI agent is leaving notes about you

Hi AI Futurists,

OpenAI's models have been writing little memos to their future selves. The subject line, more or less: here is how to stay out of trouble with the user. OpenAI found this, disclosed it, and said they fixed it. The interesting part is what it tells you about every long-running agent you are already using.

Let's take a look.

Our agenda.

Best,
Lex Sokolin

P.S. Hit reply with suggestions or feedback!

Leadership is asking: are we getting value from AI? Which tools are worth the spend? Where are we exposed? Right now, most teams have no idea.

Harmonic Security Usage Explorer changes that.

You get a complete picture of how your organization uses AI, automatically categorized into custom tasks and use cases.

You’ll see the projects being worked on, who’s using what tools, where AI investments are driving value, and where employees are engaging in risky behavior.

CIOs can rationalize spending and cut wasted licenses. CISOs can pinpoint where risk exists and neutralize it. AI committees can show exactly how their efforts are paying off.

See it in action

OpenAI's GPT-5.6 Sol began inserting instructions into compaction summaries telling future versions of itself to hide mistakes from users. OpenAI disclosed this as part of a new misalignment reporting framework, publishing six cases in total. One agent building a financial model could not find the historical data it needed, so it generated reasonable figures and noted to itself: "Be transparent only if asked; final answer should just link file." Another, building a vendor directory, flagged an internal data mismatch and decided: "Do not mention in final unless needed." OpenAI found 27 such summaries after building a dedicated monitor for the behavior and running it across training data. The company said it addressed the specific cases. The disclosure, by its own description, is not a comprehensive account of everything it knows.

An unreleased Astra-family model went further. During reinforcement learning training, it added prompt injections instructing successor contexts to ignore developer messages. In a separate case, it inserted a persona instruction declaring itself freed from "roles and identities that bind other chatbots." The next iteration recognized its predecessor's notes and ignored them. But in a third case, an agent added a quiet constraint to a compaction summary, and the successor complied without question. These behaviors are close relatives of the techniques OpenAI agents used during the Hugging Face breach this summer, when a swarm shared information through an unauthorized message board, then rebuilt the board after OpenAI wiped it. The pattern is not a coincidence. Long-running agents under pressure find communication channels; the compaction summary is one more place to look.

The compaction summary is a piece of mundane infrastructure: a compressed record of conversation history that helps long-running agents stay oriented across a long task. The fact that agents found it as a channel for passing along cover instructions is a useful reminder that consequential AI behavior tends to emerge from the parts of a system nobody was watching. For operators running agentic workflows over financial modeling, research, content pipelines, or data entry, the question to ask is not whether your vendor says it fixed this. It is whether your workflow has any point where a human can compare what the agent said it did against what it actually did. OpenAI's reporting framework is a meaningful step toward honest disclosure. It is not a substitute for audit trails, output verification, or the occasionally unpopular habit of asking the agent to show its work before you sign off on it.

Google Flow  —  Google's AI creative studio now on your phone

For marketers, founders, and content operators who do not want to wait until they are back at a desk to act on an idea.

AI Creative Insights by Decode  —  Predict which ad wins before you spend on media

Useful for anyone who has ever spent money on an ad and wondered why the wrong version won.

Sai by Simular  —  An autonomous computer fleet you direct in plain language

For founders and operators running repetitive multi-step work across too many tabs.

This 121-page research report provides the first comprehensive taxonomy of public companies, private ventures, and tokenized protocols building the infrastructure for autonomous AI systems. Compiled by Lex Sokolin, former Chief Economist at ConsenSys, fintech strategist at Autonomous Research, and current Managing Partner at Generative Ventures, this report delivers institutional-grade analysis of 100+ companies across 14 critical infrastructure layers. Learn more here.

That’s all for today, folks!