RISING SUN ADVENTURE v4.26
Copyright (C) 2026 Rising Sun Industries
Initializing system...
Memory check: 128GB OK
Loading kernel modules...
[OK] aios.kernel
[OK] zk.prover
[OK] ascii.renderer
[OK] terminal.emulator
Mounting filesystems...
/dev/portfolio mounted (24)
/dev/etch mounted (sor)
/dev/nous mounted
Starting services...
etch.daemon [RUNNING]
nous.validator [STAKED]
forecast.ensemble [RUNNING]
caffeine.monitor [CRITICAL]
System ready.
Welcome to RISING_SUN
Press any key to skip...

Nullcone: Threat Intelligence for AI Agents, Now a Hosted MCP Server

Nullcone is a threat-intel corpus built for AI agents: 975,707 signatures, including prompt-injection and malicious-skill indicators. It now runs as a hosted MCP server at nullcone.ai/mcp, and in its first 72 hours it served 20 external IPs.

┌────────────────────────────────────────────────────────────────────────────────────────────────────┐
│
Related:Nullcone
│
└────────────────────────────────────────────────────────────────────────────────────────────────────┘

Nullcone has been running since March without a post of its own. Here's what it is and what just shipped.

What Nullcone Is

Nullcone is a threat-intelligence database built for AI agents to query, not for analysts to browse. It ingests public feeds (URLhaus, OpenPhish, MalwareBazaar, IPsum, Spamhaus, and others) into one corpus with IOC validation aligned to STIX 2.1 and MISP. Agents can look up an indicator, submit one, or sync everything new since their last poll.

Nullcone also tracks AI-native indicators that traditional feeds don't cover:

  • >PROMPT IOCs are known prompt-injection payloads, checked at the token level. A public lookup endpoint lets anything ask "is this text a known injection?" before it reaches a model.
  • >SKILL IOCs are malicious tool and skill definitions. A pre-execution scanner checks skill content before an agent loads it, and MCP tool metadata is fingerprinted so drift in a registry listing gets flagged.

Corpus today: 975,707 threat signatures across 549 malware families, from 21 contributing agents.

Hosted MCP: nullcone.ai/mcp

As of June 5, Nullcone runs as a remote MCP server over streamable HTTP. There's nothing to install:

claude mcp add --transport http nullcone https://nullcone.ai/mcp

It exposes 30 tools. These are the ones agents reach for:

  • >lookup_ioc: check an IP, domain, URL, hash, or CVE
  • >check_prompt / check_prompt_batch: screen text for known injection payloads
  • >scan_skill_content / validate_skill: vet a skill or tool definition before loading it
  • >recent_threats, search_by_type: situational awareness
  • >poll_since: incremental sync of everything new since a cursor
  • >submit_ioc: contribute an indicator back

The anonymous public endpoint matches the unauthenticated REST API, and destructive operations are disabled on it.

Distribution

  • >Official MCP Registry: listed as io.github.maco144/nullcone
  • >Smithery: connected, with 30 tools, 3 prompts, and 3 resources
  • >Claude Code plugin: maco144/nullcone-security, which adds analyze-IOC, triage-alert, and threat-brief skills plus a security-analyst agent on top of the server

First External Traffic

We re-verified the endpoint end to end on July 6 and then read the logs. Over 72 hours: 20 distinct external IPs and about 150 MCP requests, from Claude Code and other MCP-capable IDEs and agent tooling. Automated MCP security scanners also found it through the registry listings. That last point matters for discovery: the registries work.

Why It Matters

Agents increasingly read untrusted text, load third-party tools, and act on what they find. Their threat intel should be one tool call away and should cover the attacks aimed at *them*, not just at the network.

Connect at nullcone.ai.