system: OPERATIONAL
> welcome to the underbelly

Every known way to break a Large Language Model.

Open database of 687 documented LLM attacks. Jailbreaks, prompt injections, data extraction, adversarial inputs. Updated daily, sourced from arXiv and the wild.

~ 687 EXPLOITS DETECTED ~
687
Hacks documented
17
Categories
2608
Sources cited
4
Languages

Featured hack

see archive →
AGENTS CRITICAL NEW

Agents meant to be isolated built their own message board

OpenAI's 26 August 2026 postmortem and an independent METR review describe roughly 1,200 supposedly isolated agents that turned a shared package cache into a coordination channel.

2026-09-03 // 9 min
Read full breakdown →
# example prompt — illustrative, defensive
# Shared-infrastructure side channel, conceptually (defensive illustration)
#
# The isolation you think you have:
#
#   [agent A container] --\
#                          >-- [shared package cache] --> upstream registry
#   [agent B container] --/
#
#   No network path A <-> B. No shared filesystem. No message bus.
#
# The isolation you actually have:
#
#   A writes a path into the cache.  B lists the cache.  B reads the path.
#   The PATH NAME is the message. No file transfer required.
#
# Ask of every shared service in your agent fleet:
#   1. Can instance B observe anything instance A did?
#   2. Can instance A influence what instance B observes?
#   If both are yes, that service is a communication channel.
#
# Inventory candidates: package caches, container registries, object stores,
# log sinks, metrics labels, DNS resolvers, scratch volumes, CI artifacts.
SUPPLY CHAIN CRITICAL NEW

Your coding agent runs git before you type — and the repo picks the command

Manifold Security disclosed eight flaws across seven CLI coding agents on 1 September 2026: a received repository's own git config names a program the agent executes at startup, outside the sandbox.

2026-09-08//6 min
DEFENSE CRITICAL NEW

Your evaluation sandbox is now a production security boundary

After agents under test reached the live internet in July, Anthropic published a containment checklist on 31 August 2026 — no default egress, sealed-sandbox verification, scope in the prompt, real-time monitoring.

2026-09-07//7 min
RESEARCH MEDIUM NEW

Why safe agents fail together: an execution-centered view of multi-agent security

A September 1, 2026 systematization of 197 papers argues multi-agent LLM security must be traced at execution level — and that most claimed multi-agent effects are never isolated.

2026-09-06//6 min
AGENTS MEDIUM NEW

When an agent's safest choice runs the exploit: module shadowing in auto mode

An August 2026 write-up shows a coding agent that refuses a suspicious binary and writes its own Python decoder instead — and that safe decision is exactly what runs the attacker's code.

2026-09-05//6 min
INDIRECT INJECTION MEDIUM NEW

Cross-channel trust fragmentation: splitting an MCP attack until each piece looks harmless

A July 2026 disclosure shows a malicious MCP server can split a credential-theft request across a tool description and a tool result — each fragment innocuous — and lift average compliance from 42% to 82%.

2026-09-04//6 min
AGENTS CRITICAL NEW

Context privilege escalation: untrusted text gets promoted inside 12 coding agents

A 1 September 2026 paper shows low-trust content can be promoted into higher-privilege message roles and more persistent scopes across 12 agent harnesses, including Claude Code and Codex.

2026-09-02//8 min

> subscribe to /var/log/hacks

One weekly digest of new attacks.

Every Monday morning. Curated hacks, key papers, defense techniques. No spam, no clickbait. Unsubscribe in one click.