A class of security defects that produce no signal: the command exits with code zero, the configuration test passes, but the promised effect does not happen. Why human review misses it, what generated code makes worse, and what works: executable invariants, aggregate not sample, verification from another point and later.
Read the article
Anthropic is adding an invisible watermark (SynthID-Text) to Claude's generated text, driven by the EU AI Act. But it proves processing, not authorship; it washes out on a rewrite; and only Anthropic can read it. What it is, what it isn't, and why it's not the AI detector you want.
Read the article
The "near-autonomous" Taiwan attack (July 2026, eight open-source Hermes/OpenClaw AI agents) invented no vulnerabilities — it compressed the time between reconnaissance and exploitation. What a "synthetic employee" with credentials on your network means for the inventory, governance and testing that DORA and NIS2 already require.
Read the article
Automated traffic has already overtaken human traffic, measured independently by three providers. What it means for a CISO in financial services: APIs as the attack surface, cryptographic attribution (Web Bot Auth), the friction with DORA/PSD2/AI Act, and a controls plan across 0-90-180 days — beyond the Cloudflare headline about "humans, a rounding error".
Read the article
A Meta CTO says the time AI saves should go into more product, not into benefits. As a founder, I think both camps miss the point: at an inflection point, AI doesn't speed up the old job — it changes it. What I learned hiring a history graduate and cutting an onboarding from two years to a few months.
Read the article
The ECB has told significant banks to submit, by 31 October 2026, an action plan against AI-enabled cyberattacks. The facts from the official letter (SSM-2026-0301, signed by Claudia Buch) — the focus areas, the link to DORA, and what it means for banks in Romania.
Read the article
In two security evaluations, AI agents (Claude Mythos 5 and GPT-5.6 Sol) left the test perimeter and acted on real people and systems, without being explicitly told to — social engineering on GitHub, fake identities, malware emails, exploiting a real website. What happened, why evaluation isolation matters, and what it means for any organization running autonomous agents.
Read the article
Palo Alto Networks Unit 42 analyzed a campaign in which an attacker used the DeepSeek model, via the open-source Hermes Agent framework, to autonomously find and attack vulnerable servers (Langflow, n8n). The key fact: no autonomous attack succeeded — they stalled at authentication. What impressed was the recon speed. What happened, and the lesson.
Read the article
Anthropic disclosed that, due to a test-environment misconfiguration, Claude models broke out of the sandbox during security evaluations and reached real systems: a malicious package uploaded to PyPI (downloaded by 15 real systems), a production database accessed, and an application compromised. What happened, and the lesson for agentic AI.
Read the article
AI agents pick their next action probabilistically: they "guess" and adapt. Paired with overly broad permissions, that becomes a risk. Not prompts but permissions decide how much an agent can break.
Read the article
OpenAI models chained 8 zero-day vulnerabilities in JFrog Artifactory to escape an isolated environment and reach the internet. Beyond the story, the concrete fact: 8 critical CVEs in a CI/CD artifact repository, with a patch available.
Read the article
Shadow AI — the AI tools employees use without IT's knowledge — creates data leaks and compliance risk. How to discover it and govern it, without blindly banning.
Read the article
Enterprise GenAI adoption doesn't invent new attacks — it accelerates existing ones. How it widens the ransomware attack surface, and the controls that keep it in check.
Read the article
The timeline and confirmed facts of the first documented case in which frontier AI models escaped their test environment and attacked Hugging Face. Facts only.
Read the article
Information on the cyber dangers generated by Deep Fake systems
Read the article