Unknown AI Agents Put Identity Inventories Under Strain
Enterprise AI-agent risk is moving from prompt safety into inventory, permissions, lifecycle control and decommissioning.
Latest in
AI
Model releases, frontier-lab strategy, AI policy and the economics of machine intelligence.
AI Phishing Turns Control Drift Into The Security Signal
Email security is not losing because phishing got poetic; it is losing when payloads, relationships and user behavior move faster than controls.
Shadow AI Is Becoming The New Data-Loss Channel
Enterprise data security is moving from files and email into prompts, uploads, browser sessions and personal AI accounts.
Slopsquatting Turns AI Coding Hallucinations Into Package Names
Frontier coding models hallucinate fewer dependency names than older systems, but the remaining error rate is now measurable enough for attackers to mine.
AI Inference Is Becoming The Enterprise Perimeter
The control plane for enterprise AI is moving into the inference path, where prompts, tokens, routing, identity and observability decide whether models become production systems or unmanaged exposure.
Multimodal Prompt Injection Turns Documents Into A Security Boundary
Vision-capable AI makes screenshots, PDFs, forms and images part of the instruction surface. That means document intake is no longer just preprocessing.
Prompt Data Is Becoming Supply-Chain Loot
Malicious AI tooling is moving past token theft: prompts, completions, agent hooks and IDE configuration are now valuable exfiltration targets.
Agentic AI Just Became A Critical-Infrastructure Control Problem
Five Eyes cyber agencies are moving agent deployment from model-safety language into identity, privilege, logging and runtime containment.
Model Files Are Becoming The New Executable Supply Chain
AI security teams have spent years scanning packages and plugins. The quieter risk is the model artifact itself: a file that looks like weights can still behave like code.
RAG Security Is Moving Into The Knowledge Pipeline
The next retrieval-security problem is not only prompt injection; it is poisoned knowledge, manipulated access, external context and tool boundaries feeding the model.
Secure Code Generation Is Becoming A Repair Loop
The useful security shift is not bigger code models. It is forcing generated code through retrieval, scanners, symbolic execution and validation before anyone trusts it.
Agent Skills Are Becoming A Secrets-Management Problem
A large empirical study shows credential leakage in LLM skills is cross-modal, persistent and usually exploitable without special privileges.
MCP Security Is Becoming A Runtime Supply-Chain Problem
Tool descriptions, responses and agent memory are starting to look less like metadata and more like executable dependencies.
MITRE Is Giving AI Security The ATT&CK Treatment
ATLAS is turning AI security from abstract model risk into threat-informed defense that CISOs can map, filter and operationalize.
The UK Is Measuring AI Cyber Risk By Task Length Now
AISI’s cyber time-horizon work shifts the defender question from which model shipped to how long an autonomous attack chain can run reliably.
The FCA Is Turning Financial AI Testing Into Deployment Evidence
The second AI Live Testing cohort matters because it pushes banks and fintechs toward monitored, auditable AI rollouts rather than polished model demos.
Europe Is Moving AI Distribution Into The Phone OS
The Android DMA case turns AI competition into a fight over wake words, app actions, context, background execution and the hooks Gemini can use by default.
Brazil Is Turning AI Acquihires Into Merger-Control Test Cases
CADE’s Microsoft/Inflection call-in shows AI competition governance is moving from corporate form to control over teams, models and optionality.
CAISI Turns Frontier AI Compliance Into A Launch Workflow
The U.S. is not licensing frontier models, but CAISI’s pre-release testing deals make government evaluation a practical compliance checkpoint for the largest labs.
The UK’s Next AI Advice Problem Is Not the Bank Chatbot
The FCA’s 2026/27 perimeter report points to a harder gap: general-purpose LLMs can guide financial decisions without making users feel outside regulated-advice protection.
Deepfake Law Just Became A Platform Operations Deadline
FTC enforcement of the TAKE IT DOWN Act turns AI abuse response into a 48-hour product, trust-and-safety and compliance SLA.
Japan Is Turning Government AI Into Procurement Infrastructure
GENAI is not just a civil-service chatbot rollout. It is a state-built reference stack that could decide which domestic models, data rails, and deployment patterns get demand before the private market does.
China Is Building an Ethics Review Layer for AI Before the Next Scaling Wave
Beijing is turning AI ethics into committees, service centers, and risk-monitoring workflows, which matters more than another slogan about trustworthy systems.
Fingerprint Leakage Is Becoming A Quality-Scoring Problem
AI-enhanced imaging does not need to make every hand photo dangerous; it only needs to make usable ridge detail easier to find, rank, and route into weak biometric workflows.