Reading List - September 22, 2025
AI Agent vulnerabilities, AI economics, and some attention deep dives - we've had a fun week of reading!
OpenAI Burns the Boats
By: Ethan Ding
What is OpenAI's plan to beat Anthropic in the AI race? Check it out!
The Economic Impacts of AI: A Multidisciplinary, Multibook Review
By: Kevin A. Bryan
This essay reviews seven books from the past dozen years by social scientists
examining AI’s economic impact. The author gleans interesting insights but, like all of us, can't answer what's next for the labor market in the age of AI.
What are you Sinking? A geometric approach on attention sink
By: Ruscio et. Al
This paper shows that “attention sinks” are effective reference points that help tokens relate to each other, and the type of positional encoding changes how these reference points form.
Attention Sinks from the Graph Perspective
By: The Tensor Throne
This one's a little math-heavy, the author breaks down possible mechanisms that may bias Causal Transformers to accumulate attention on its first few tokens.
AGI and the Political Order
By: Gustavs Zilgalvis
Not all of us are convinced that AGI will arrive, but if it does, the author has an interesting perspective on what that future holds for social structures.
Making the Case for a Third AI Technology Stack
By: Brooke Tanner and Andrew W. Wyckoff
These researchers from the Brookings Institute argue that Europe and other democratic countries need to develop their own AI infrastructure to ensure values like transparency, fairness, and accountability (because the US and China won't).
Mind the Gap: Time-of-Check to Time-of-Use Vulnerabilities in LLM-Enabled Agents
By: Derek Lilienthal and Sanghyun Hong
This is a cool academic paper from researchers at OSU, who've developed a new system for benchmarking and testing prompt injection vulnerability of LLMs.
The Hidden Risk in Notion 3.0 AI Agents: Web Search Tool Abuse for Data Exfiltration
By: Abi Raghuram
Everybody's spinning up their own MCP/AI Agent platforms, and Notion (like many in the space) prioritized speed over security. Not great!
Tweet of the Week
We mourn the co-option of em-dashes by AI - we've got some heavy users at Confident Security. Maybe the triple em-dash is the way forward—we're open to trying it.
