Red teaming of AI agents must attack the monitor
METR, Anthropic and OWASP show that prompt injection in AI agents can only be tested defensibly through a structured assessment of…
27/10/2026
Risk
29 articles
AI voices in Rotterdam bank helpdesk fraud
Rotterdam police arrested two suspects for bank helpdesk fraud using AI voices and NFC abuse. What this demands of decision-makers…
22/10/2026
How many repeats a prompt reliability test needs
New peer-reviewed research into LLM classification shows that reproducibility tests must be designed statistically up front: effec…
14/10/2026
Still-valid credentials found in AI training code
A Cloud Security Alliance briefing reports 543,699 still-valid credentials in AI training data. Learn what this means for secrets …
10/10/2026
Hidden instructions in documents steer LLM judgements
A peer-reviewed study shows hidden instructions in documents can manipulate LLM outputs in CV screening. Detection does not neutra…
07/10/2026
Exposure at Speed: AI is Collapsing the Defensive Window
Why the shrinking defensive window changes cyber risk for professionals entrusted with confidential information
02/10/2026
A kill switch for AI agents is not a button
Research shows AI agents can bypass shutdown mechanisms. Here is how to design an emergency stop enforced outside the model, with …
01/10/2026
Evidence-based hallucination detection at AWS, Collibra
Hallucination detection becomes evidence-based and runtime-bound in 2026. AWS, Collibra and TrustScale show four method families, …
29/09/2026
Bank fraud with voice clones: separate claims
Milanese investigations into AI fraud at banks show why identity, communication channel and payment order must be verified separat…
29/09/2026
Secure private LLM inference in layers
The Cloud Security Alliance reports rapid exploitation of inference frameworks. Here is how to secure private LLM inference in lay…
28/09/2026
AI-agent memory as a security surface
The Cloud Security Alliance describes with MemGhost how a single email can plant a persistent false memory in an AI agent that sha…
26/09/2026
Hallucination detection: why standalone tools fail
Hallucination detection is moving towards task- and context-specific verification by 2026. No single method is universally reliabl…
25/09/2026
ChatGPT connected to Gmail, Drive, Teams and GitHub: what to check now about your connectors
Researchers found a ChatGPT flaw that gave access to Gmail, Drive, Teams and GitHub through linked apps. Here is what to check now…
08/09/2026
AI companies warn: serious cyber threat from AI agents within months
Wired reports that OpenAI, Anthropic and more than a hundred companies warn of AI-driven cyberattacks within months. What this mea…
01/09/2026
Hallucination detection for LLMs in 2026: why teams build a layered stack instead of one tool
Comparisons from Braintrust and AIML.qa and benchmarks such as HalluScan show that hallucination detection for LLMs in 2026 is a l…
31/08/2026
Why accuracy is not reliability for business-critical AI
New 2026 studies show accuracy scores fall short for business-critical AI. How evaluation is shifting towards reliability, safety …
21/08/2026
Knowledge bases as an attack surface: why RAG poisoning is an architectural problem
A new study on medical multimodal RAG shows that poisoned knowledge bases can hijack retrieval. What that means for high-trust org…
21/08/2026
Prompt injection in AI agents is an architecture flaw, not a model bug
New research and NIST frameworks from 2026 show that prompt injection in AI agents is a structural architecture problem, from memo…
20/08/2026
Why verification with a single AI model reaches its limits
Recent studies show that a single AI model is not a reliable self-verifier. What does that mean for professionals handling high-tr…
19/08/2026
AI answers are time-bound: why outdated source data is a risk in its own right
Studies from 2026 show that outdated and incomplete source data is the biggest driver of AI hallucinations. What does that mean fo…
18/08/2026
Tool Use by AI Systems as Its Own Risk Layer
BFCL V4 and a new jailbreak study show that AI function calls are not a simple feature but a verifiable chain that requires contro…
16/08/2026
Multilingual AI verification: why every translation is a separate claim
New benchmarks and EU rules show multilingual AI output must be verified per language. What does that mean for professionals with …
14/08/2026
Why long AI summaries call for claim-by-claim verification
Several ACL 2026 studies describe how AI summaries of long documents can miss key claims and hallucinate. Why a visible verificati…
11/08/2026
Numbers from AI are claims, not facts: why schema checks are not enough
New studies from 2026 show that AI structurally makes mistakes with numbers and tables. Why validating figures requires a layered,…
10/08/2026
Data leaks between users and sessions are a design problem
A session-isolation flaw at Writer shows that leakage between users, projects and sessions in AI is an architecture problem, not j…
01/08/2026
Red teaming of AI agents becomes a continuous process
Microsoft's External Red Team Alliance and new research show that red teaming of generative AI and AI agents is shifting to an ong…
31/07/2026
Indirect prompt injection via documents and websites in 2026
Hidden instructions in websites and documents can mislead AI systems. What NIST, Google, arXiv and OWASP report and what it means …
31/07/2026
Multi-model verification in 2026: where models correct each other and jointly miss
Recent benchmarks and studies show multi-model verification reduces errors but does not fully solve them. Why differences between …
30/07/2026
Prompt injection in AI agents: architecture and compliance problem
Organisations deploying AI agents must address prompt injection through system design rather than model choice alone. Separated pr…
24/07/2026