MACHINE LEARNING
AI
Prompt injection, deepfakes, agentic AI and MCP, security and policy at the frontier.
Where artificial intelligence meets the threat model: prompt injection, agentic and MCP attack surfaces, deepfake-driven social engineering, shadow AI, and the governance rules catching up with all of it. Start with the prompt-injection defender’s playbook, the MCP attack surface and the first AI agent to run a ransomware attack.
Start here
- Prompt injection: the defender’s playbook
- MCP security: the attack surface mapped
- JadePuffer: the first AI agent to run a ransomware attack
- Deepfake vishing: voice-clone fraud explained
Latest AI security coverage
- Stolen AI prompts: three providers breached, Dragos findsDragos hid trackable secrets in two million AI prompts. They were stolen from three breached providers, read by resellers and fetched despite privacy flags.
- ARTEX: the AI hacking tool behind the FBI warningARTEX is an AI penetration testing tool the FBI says attackers have used since August 2026. Who used it, how they got in, and what security teams should fix.
- 36,769 exposed AI endpoints, only 2% ask for a loginMysterium VPN counted 36,769 self-hosted AI endpoints reachable on the public internet. Only 2.02% return an authentication challenge, and that is a floor.
- What is slopsquatting? AI package hallucinations, explainedSlopsquatting is a supply-chain attack that weaponises the package names AI coding assistants invent. How it works, what the research measured, the real 2026 incidents, and how to defend an AI-assisted pipeline.
- “Delve” is dead: AI writing tells expire in 18 monthsThe word delve fell 94% from its 2024 peak while dash use doubled. We measured 21,442 arXiv abstracts: AI writing tells expire in about 18 months.
- Vibe coding is shipping vulnerabilities at scale in 2026AI-generated code is fast and insecure. Veracode found 45% of samples carry OWASP Top 10 flaws, roughly a fifth of AI-suggested packages are hallucinated, and vibe-coded apps are already leaking real data.
- Prompt injection left the lab in 2026. It is in the wild nowPrompt injection stopped being a lab demo in 2026. A one-click Claude Desktop flaw, AI browsers leaking credentials, and agents tricked into crypto payments show the attack is now operational, not theoretical.
- Shadow AI is the new stealer-log jackpot in 2026Shadow AI is now a top insider threat: 45% of staff use AI on corporate devices, most via personal accounts. When those sessions land in stealer logs, attackers inherit the AI and everything typed into it.
- macOS.Gaslight: malware that prompt-injects your SOCSentinelLABS found macOS.Gaslight, a DPRK-linked Rust backdoor that embeds 38 fake system messages to make AI malware-triage agents abort. It is the first implant built to attack the analyst’s LLM, not the sandbox.
- JadePuffer: the first AI agent to run a ransomware attackSysdig documented JADEPUFFER, the first ransomware attack whose technical execution was run end to end by an AI agent. A human handler still picked the target, built the infrastructure, and supplied credentials.
- Agentic AI threats: how MCP becomes an attack chainAgentic AI moves the threat from what a model says to what it does. We map how MCP turns goal hijacking, tool misuse, and privilege abuse into a working attack chain, and the controls that contain it.
- MCP security in 2026: the attack surface mappedA technical map of the Model Context Protocol attack surface in 2026: tool poisoning, line jumping, rug pulls, tool-chaining exfiltration, token sprawl, and the RCE flaws that turned MCP servers into entry points.
- Build a secure MCP server in 2026: a hardening guideA practitioner’s hardening guide for MCP servers in 2026: OAuth 2.1 auth, least-privilege tool scopes, sandboxing, egress control, and the tool-definition pinning that blocks poisoning and rug pulls.
- Deepfake vishing 2026: voice-clone fraud explainedDeepfake vishing uses AI-cloned voices to defeat phone-based trust. Here is how voice-clone fraud works in 2026, what it costs, and how to stop it.
- Fable 5 and Mythos 5: the US-only gate threat actors will beatWashington cut Fable 5 and Mythos 5 for every foreigner. Meanwhile 34,814 stolen Claude sessions and 74,114 ChatGPT cookies already sit in infostealer logs. A citizenship gate is just KYC, and KYC loses.
- Prompt injection: the 2026 LLM defender’s playbookPrompt injection is now the dominant attack vector against LLM-powered applications, and most teams shipping AI features don’t have a defensive playbook. We map the attack taxonomy, walk through real exploit patterns, and lay out the controls that actually contain the blast radius.
- How large language models are reshaping phishingPre-LLM phishing was constrained by language. Post-LLM phishing is not. The result is a measurable upgrade in lure quality, a wider reach into non-English-speaking markets, and an emerging class of personalised attacks that were previously economically unviable.
Every AI article
Every article we have published in this section, newest first.
- 36,769 exposed AI endpoints, only 2% ask for a login
- What is slopsquatting? AI package hallucinations, explained
- “Delve” is dead: AI writing tells expire in 18 months
- Vibe coding is shipping vulnerabilities at scale in 2026
- Prompt injection left the lab in 2026. It is in the wild now
- Shadow AI is the new stealer-log jackpot in 2026
- macOS.Gaslight: malware that prompt-injects your SOC
- JadePuffer: the first AI agent to run a ransomware attack
- Agentic AI threats: how MCP becomes an attack chain
- MCP security in 2026: the attack surface mapped
- Build a secure MCP server in 2026: a hardening guide
- Deepfake vishing 2026: voice-clone fraud explained
- Fable 5 and Mythos 5: the US-only gate threat actors will beat
- Prompt injection: the 2026 LLM defender’s playbook
- How large language models are reshaping phishing
















