RAG
retrieval-augmented generation
Retrieval-augmented generation: an AI model is handed relevant documents at question time so it answers from your data instead of memory.
In our briefings
The claim each of the newest briefings on RAG would have you leave with.
Woven by Toyota's AI agent sits at a new third step of the company's security-exception review, ahead of the human reviewer. Juan Corena says the outputs include a recommendation to approve, a recommendation to reject, or a request for more information with questions that may matter for granting it, and once the agent judges it has enough information the ticket goes to a human. His reason: "eventually a human needs to take the responsibility and understand all the consequences of what is being done there." He adds that simply asking whether a request complies with the standards would always get "no", since that is why it was requested, so the design had to allow for it.
Woven by Toyota's exception agent recommends; a person decides, because a human has to answer for itBI-FIND-2026-2242
NatWest's complaints agent writes investigation summaries for complaint handlers and has been live since February. Around 14,000 of its cases were scored by both a human validator and an LLM judge. Osborne says the judge's faithfulness scores averaged 0.97 with very little spread, against a human mean of 0.83 with a much wider one. The average error of 0.2 might look acceptable, he says, "but the correlation is almost zero." Changing only the model behind the judge cut most of its margin over the pass threshold, which is why he says the judge model "becomes part of your configuration". NatWest moved the judge to a different model family from the agent, which he presents as the effective way to bring the two scores closer together.
NatWest found its LLM judge barely tracked human reviewers, and an 83% pass rate hid 43% on topic changesBI-FIND-2026-2060
McMahon says Barclays spent the past six to nine months moving from "pockets of innovation" to one control plane, with the rule that "everything has to come through the single gateway", is registered as an enterprise artifact with its permissions and claims, and runs only in approved runtimes. Budget controls keyed to the requesting identity already sit at the gateway and in the bank's own agent SDKs, which he says is "sort of" what the AAuth protocol's budgets would do. He puts that down to size: "we can do that because we are one of those fishbowls I mentioned", a finite organisation that can force all traffic through the gateway. On the open web, he asks who runs that gateway, and answers "No one." He gives no numbers or deployment counts, and names no gateway product the bank runs. He calls it scaffolding that the bank is almost ready to test at massive scale, and says intent-based authorization is still being tried out rather than run for real.
Barclays mandates one gateway for its agents' budgets, blocks and audit, and its AI engineering director says the setup is not yet tested at scaleBI-FIND-2026-1478
›Cited in 14 briefings
Every briefing that discusses this term, newest first.
- Woven by Toyota's exception agent recommends; a person decides, because a human has to answer for it
Video · · Agentic AI Foundation · Uploaded November 4, 2026
- NatWest found its LLM judge barely tracked human reviewers, and an 83% pass rate hid 43% on topic changes
Video · · AI Engineer · Uploaded November 4, 2026
- Barclays mandates one gateway for its agents' budgets, blocks and audit, and its AI engineering director says the setup is not yet tested at scale
Podcast · · Insecure Agents · The Fishbowl Problem: Agent Identity Across Clouds and Vendors, with Andy McMahon (Barclays) · Oct 5, 2026
- Anthropic asks for per-agent identity; its guide leaves Claude Code sub-agent credentials unspecified
Article · · Anthropic (Claude) · Anthropic · May 27, 2026
- OWASP's agent Top 10 blames identity systems built for people, and does not say how it chose or ranked its risks
Article · · OWASP GenAI Security Project · John Sotiropoulos et al. · Dec 9, 2025
- NIST's agent identity work starts from OAuth, SPIFFE and SCIM and defers agents from outside the enterprise
Article · · NIST National Cybersecurity Center of Excellence · Harold Booth et al. · Feb 5, 2026
- NIST's request on securing AI agents asks about hijacking and rollback, and never about who the agent is
Article · · Federal Register (govinfo) · NIST Center for AI Standards et al. · Jan 8, 2026
- An MCP threat survey shows agents trusting tool names and descriptions, and skips the spec's own OAuth rules
Paper · · arXiv · Xinyi Hou et al. · Oct 7, 2025
- LPCI paper names instructions planted in agent memory that fire in a later session; its 43% failure rate matches a five-model average
Paper · · arXiv · Hammad Atta et al. · Jul 14, 2025
- AgentDNS gives AI agents one sign-in by having a central root hold each vendor's key
Paper · · arXiv · Enfang Cui et al. · May 28, 2025
- OWASP's multi-agent threat guide finds the hop where a user's permission check drops out, but names no fix
Article · · OWASP GenAI Security Project · OWASP Agentic Security Initiative · Apr 22, 2025
- In a lab test with simulated hops, one poisoned email spread through replies that an AI assistant drafted
Paper · · arXiv · Stav Cohen et al. · Jan 30, 2025
- AgentPoison's few poisoned memory entries steered agents 62.6% of the time on average, not the abstract's 80%
Paper · · arXiv · Zhaorun Chen et al. · Jul 17, 2024
- In 2025, Kuppinger said Microsoft was underestimated and expected EU wallets to make verification reusable
Podcast · · Identity at the Center · #326 - IAM in 2025 with Martin Kuppinger · Jan 20, 2025