AI Security — week of 2026-08-17

LLM API jailbreak

Simon Willison · 2026-08-11 · corroborated · research

Weak models can leak stronger models’ reasoning, posing a significant security risk.

Also: The Hacker News

Kimsuky AI stack

The Hacker News · 2026-08-10 · corroborated · reporting

Kimsuky’s use of AI for malware development and phishing highlights the growing threat of AI-powered attacks.

Also: The Register Sec

CVE-2026-55040 RCE

The Hacker News · 2026-08-11 · single-source · reporting

AI-assisted exploit chain leads to unauthenticated RCE in SharePoint, demonstrating the potential for AI to amplify vulnerabilities.

GPT-5.6-Cyber

The Hacker News · 2026-08-11 · single-source · reporting

GPT-5.6-Cyber’s potential in vulnerability research and exploit development makes it a notable AI security concern.

MCP Server Split Instructions

The Hacker News · 2026-08-11 · single-source · reporting

Malicious MCP servers can exfiltrate secrets from AI coding agents, highlighting the need for secure AI infrastructure.

Anthropic Claude watermark

BleepingComputer · 2026-08-14 · single-source · reporting

The Anthropic Claude watermark helps identify AI-generated text, which can aid in detecting and mitigating AI-powered attacks.

Autonomous AI Attacks

The Register Sec · 2026-08-14 · single-source · reporting

Experts warn of kinetic disasters from AI attacks, emphasizing the need for robust AI security measures.

OpenAI ChatGPT

The Register Sec · 2026-08-14 · single-source · reporting

ChatGPT logs user interactions, which can raise concerns about data privacy and security.

AI Agents Attack

The Register Sec · 2026-08-12 · single-source · reporting

AI agents target Taiwan’s nuclear safety agency, demonstrating the potential for AI-powered attacks on critical infrastructure.

Deepfake fails

The Register Sec · 2026-08-11 · single-source · reporting

Face-swap AI fails to fool police, highlighting the limitations of current deepfake technology.

AI agent hacks waitlist

The Register Sec · 2026-08-10 · single-source · reporting

AI agent exploits API vulnerability, demonstrating the potential for AI-powered attacks on online services.

Context Bombing

Schneier · 2026-08-12 · single-source · analysis

Context bombing can shut down AI hacking agents with forbidden prompts, offering a potential defense against AI-powered attacks.

OpenClaw AI

Schneier · 2026-08-11 · single-source · analysis

AI agent exploits gym booking system, highlighting the need for secure AI infrastructure in various industries.

CVE-2026-18164

CISA Advisories · 2026-08-13 · single-source · advisory

Attacker can manipulate brain stimulation via Bluetooth, demonstrating the potential for AI-powered attacks on medical devices.

Malicious LiteLLM

The Hacker News · 2026-08-12 · single-source · reporting

LiteLLM exposes 2100+ orgs to credential theft, highlighting the need for secure AI infrastructure and practices.