Research
We publish two kinds of writing. The explainers set out the protocols identity runs on. The research desk reads what the industry says about AI and identity, checks it against the primary sources, and keeps score on its own forecasts.
IAM Insights
From the research desk
The newest briefings from IAM Insights, the firm's public research desk.
The desk’s ledger, as of
- Forecasts
Every forward-looking call IAM Insights publishes, held to its expiry — and settled publicly when the window closes.
- Research agenda
Every open question IAM Insights has published at the intersection of cybersecurity and AI — why each matters, where we'd start, and which later briefing answered it.
- Independence
Who pays for IAM Insights, which identity platforms Boston Identity implements for clients, and what the desk reads.
Knowing who an agent is and what it may reach is not enough, six voices say: check each action as it runs
Thirteen podcast episodes and talks from December 2025 to November 2026 ask how far an AI agent can be trusted. Six voices, among them Ciena, JPMorganChase, Speakeasy and Radware, say an agent's identity and permissions do not show whether its next action is right. Others want every agent traced to a person or company that answers for it, or kept in a watched sandbox, and the voices differ on how much an agent's intent can carry.
MCP-Guard's classifier alone came within half a point of its best full pipelines' F1 on the authors' test set, and ran faster
Wenpeng Xing and colleagues at Zhejiang University, the Chinese University of Hong Kong and Shandong University built MCP-Guard, a proxy that screens traffic between AI models and their Model Context Protocol (MCP) tools with pattern rules, a fine-tuned text classifier and then an LLM judge. On their own benchmark it beat three open-source MCP scanners on F1, a score that balances missed attacks against false alarms. Run alone, the classifier came within half a point of the best full pipelines' F1 and was faster than every one of them.
A new STIR draft would let a number's holder require AI agents calling from it to say so
Jon Peterson of TransUnion, an author of several STIR standards the draft builds on, wants the signatures that vouch for a caller's number to cover AI agents placing or answering calls for someone. His -00 individual draft adds an "act" claim naming the agent and an "intent" claim that carries, or links to, a separately signed record of what the person or company authorized. It concedes that the step turning a request into that record, typically done by a large language model, cannot be mechanically verified.
In an a16z survey of 100 CISOs, agent security climbs the 2027 list and access management slips
Andreessen Horowitz surveyed 100 CISOs at companies with more than 1,000 employees. Internal AI agent security gained the most ground as a 2027 priority, and 45% plan to buy or upgrade it within a year, while fewer put access management or privileged access management in their top five than they recall doing for 2026. The firm discloses stakes in two of the startups it names and does not publish its sampling method or full questionnaire.
Palo Alto's 109-to-1 machine identity ratio is mostly AI agents, and cannot be read as last year's 82 grown
Palo Alto Networks' 2026 Identity Security Landscape says organisations average 109 machine identities for every human, and that 79 of those per human are AI agent identities. Its public excerpt names no survey firm, question or counting unit for either number and prints no earlier ratio, so a reader has no way to line the 109 up with the 82 to 1 of CyberArk's 2025 edition or with Entro's 144 to 1.
A 2023 paper on agentic AI harms argued that calling a system an agent leaves the responsibility with people
Alan Chan and 21 co-authors from Mila, Cambridge and other academic groups wrote this 2023 position paper, which OpenAI's agent-governance paper later cited for the term "agentic AI systems". It treats agency as something a system has more or less of, casts the human as principal and the system as agent, and holds that the label moves no moral responsibility off people. Its main text never mentions identity, credentials or access control.
Explainers
Protocol explainers in plain language, in four series.
IAM Handbook
The firm's handbook on identity and access management, free to download.
Download · PDF · 227 pages · 10 MB