FTC
Federal Trade Commission
The US consumer-protection regulator that tracks reported fraud losses (its Consumer Sentinel Network) and enforces rules against deceptive and impersonation-based scams.
In our briefings
The claim each of the newest briefings on FTC would have you leave with.
Asked which AI fraud risk gets too little attention, Larsen says the central fraud question "might shift" from whether this is really the customer transacting to whether it is the customer's agent, and what that agent is "actually authorized to do". Payments and digital identity systems, she says, have "generally assumed" a human behind each transaction; agents, she says, raise the question of how to authenticate delegated authority and scope as well as identity, down to whether an agent had permission to spend this much, on this, right now. Among the gaps she names in the law is one we read as the same problem from the other side: legal architecture is "often better at punishing someone after fraud has occurred than establishing the trustworthy identity before a transaction".
The fraud question may shift to: is this your agent, and what is it authorized to do?BI-FIND-2026-0119
Mireshghallah and Li had GPT-4.1 sort a decade of privacy papers from the main machine learning, language and security conferences into five kinds of incident. Training-data leakage took 48.4% and direct chat leakage through the provider 43.6%, together 92%. That first bucket does not separate pre-training from post-training work, though the authors themselves call verbatim regurgitation of pre-training data overstated and memorization in fine-tuning and post-training a real, understudied risk. Inferring personal attributes took 5.8%, context leakage through agents, retrieval and prompt injection 2.0% (25 papers), and aggregating personal data from public sources 0.2% (3 papers). The split depends on a topic-to-incident mapping the authors wrote, and the classifier was checked on 50-paper samples at 96% to 100% accuracy. The counts in their charts add to 1,261 papers, not the 1,322 the text reports, and the paper does not explain the gap.
A decade of AI privacy papers: nearly half on training-data leakage, 2% on agents leaking user contextBI-FIND-2025-0181
Cited in 2 briefings
Every briefing that discusses this term, newest first.
- The fraud question may shift to: is this your agent, and what is it authorized to do?
Podcast · · AI Security and the Law · Blurred Lines: How AI Is Rewriting the Rules of Fraud, Identity, and Defense · Aug 21, 2026
- A decade of AI privacy papers: nearly half on training-data leakage, 2% on agents leaking user context
Paper · · arXiv · Niloofar Mireshghallah et al. · Oct 2, 2025