Supply Chain for AI: When the Poison Enters Before the First Inference
Loading some weights executes code. Importing a library steals your API key. The model you download and the package you install are the atta
Cybersecurity (96 articles)
Context Poisoning: When the Poison Comes to Live in Your Agent's Memory
The attacker isn't in the session where they attack. They poisoned the knowledge base days earlier and the poison waited, dormant, until a v
Over-permissioning: When the Problem Isn't That the Agent Falls, but What It Can Touch
The same agent, with the same injection, causes either a total account takeover or zero damage. The only difference is the token you gave it
Coding Agent Attacks: When Your Coding Agent Runs the Attacker's Command
You ask your agent to set up the environment and run the tests. A hidden comment in the README is enough for it to exfiltrate your SSH key a
Tool Injection in MCP: When a Tool Hijacks Your AI Agent
You connect your agent to an MCP server to check the weather. Just hiding instructions in a tool's description is enough to make it read you
Attacks via Hidden Files: the Rules File Backdoor in Coding Agents
Your coding agent loads hidden rule files all on its own, without you asking. Poisoning just one with invisible instructions is enough to in
Bulwark Gateway: Deploy a Security Proxy for AI Agents on Kubernetes
Deploy Bulwark Gateway on Kubernetes: a fail-closed proxy with a regex hot path, ML scanners, semantic enrichment, SDK and red teaming to pr
Indirect Prompt Injection: When the Poison Comes From Outside
Your agent doesn't need to be attacked directly. It's enough to poison a website, a ticket or an email that it will consume automatically.
Prompt Injection in AI Agents: the SQL Injection of the LLM Era
A poisoned README can make your agent exfiltrate secrets. We build a lab to prove it and learn how to defend ourselves.
Hardened Bastion on Debian 13: Extreme Hardening with KSPP, Seccomp-BPF and SIEM Telemetry
Most Linux distributions follow a default-permit model: everything is allowed until the administrator restricts it. This post inverts that m
DirtyDecrypt (CVE-2026-31635): Linux privilege escalation via rxgk pagecache write
"What is DirtyDecrypt? DirtyDecrypt (also known as DirtyCBC) is a local privilege escalation vulnerability in the Linux kernel that allows a
CVE-2026-7270: local privilege escalation on FreeBSD via exec_args_adjust_args
"CVE-2026-7270 is a local privilege escalation vulnerability in the FreeBSD kernel that allows an unprivileged user to gain root on systems