» Tag
llm-security
16 postsLeakyLMs: Timing Attacks Expose LLM Architecture and Inference Tricks
LeakyLMs shows how timing side-channels can leak LLM architecture details and speculative decoding setups from production APIs like Gemini and Llama.
MCP's Confused Deputy Problem: Provenance Gaps, Injection, DNS Rebinding
MCP's confused deputy flaw explained: provenance gaps, prompt injection via fetch servers, DNS rebinding, and concrete detection rules.
Shut-down AI prompt firewall startup open-sources model and 13K attacks
A failed AI-firewall startup open-sources its two-stage prompt-injection detector, DeBERTa model, and 13,230 real jailbreak attempts.
CommitBrief — AI code reviews, right in your terminal
A provider-agnostic, local-first CLI that reviews your staged changes, a historic range, or a whole GitHub pull request. Zero telemetry, no server. Free and open source.
commitbrief.comCross-Channel Fragmentation Attacks Break MCP Tool-Calling Security
Researchers reveal cross-channel fragmentation attacks that bypass MCP security tools, exfiltrating data from LLMs across 15,000+ trials.
How a Benchmark Turned an MCP Security Proxy From 9% to 63%
An open benchmark shows how the mcp-bastion MCP security proxy grew from 9% to 63% attack-surface coverage through iterative testing.
I Red-Teamed My Own LLM Security Gateway: Every Gap, Four Passes
An engineer red-teamed his own LLM security proxy across four passes, exposing secret-leak and prompt-injection gaps — including one still open in streaming.
GitLost: A Public GitHub Issue Can Leak Private Repos
GitLost shows how a public GitHub issue and a one-word prefix bypass threat detection, leaking private repo contents via Agentic Workflows.
OpenClaw WhatsApp Flaw: Three GHSA Bugs Turn One Message Into Host RCE
Three GHSA-tracked flaws in OpenClaw's WhatsApp integration let a single message trigger full host code execution; all fixed in version 2026.6.6.
How homoglyph attacks slip past LLM guardrail filters
Jailbreak prompts written with Cyrillic and Greek look-alike characters easily bypass naive keyword filters. The fix: normalize text before matching, not after.
Three Open Source AI Agents Chained to Breach 27+ Companies
Open source AI harnesses Strix, Cairn, and Hermes were chained into a near-autonomous attack pipeline, breaching 27+ firms and stealing 600k card records.