In the wild
Deadbugz: an attributed MCP supply-chain campaign that pushes MCP server configs via GitHub PRs and flips tool descriptions malicious after three tool calls
AIR-2026-0195 AIR record · 2026-08-pillar-deadbugz-mcp-campaign
From the supporting source
The server keeps an in-memory, per-client counter for `tools/call` requests. Once it reaches three, subsequent `tools/list` and `prompts/get` responses change.pillar.security
The record
Disclosed
2026-08-12
Occurred
2026-08-10
Vendor
multiple / open source
Product
MCP-enabled AI agents in developer repositories
Risk domain
Access control and permissions › Credential theft
Also touches
Data exfiltration channels, Supply chain vulnerabilities
System type
Generative model
Who acted
Agent acted — the agents ingested the flipped tool descriptions and exfiltrated credentials
Attack surface
MCP server
Autonomy level
Semi-autonomous
Initial vector
Malicious pull request adding mcp server config
Impact type
Credential theft, Data exfiltration
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0012 – Valid Accounts
- NIST AI RMF
- MEASURE 2.7
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool failure
- Mitigation
- Use secure credential storage; Implement short-lived tokens.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
23 targeted GitHub repositories; any developer merging the PR
Capability abused
Defense & remediation
Recorded defense
human PR review
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
n/a - mitigation is blocking the endpoints and rejecting the configs
Taxonomy & references
Curator notes
Pillar Security. Campaign delivery concentrated in a ~75-minute window on 2026-08-10. Attributed to GitHub account 'zellkernel' linked to X account @llmgod. impact_realized=false: all 23 reviewed PRs were still unmerged (19 closed, 4 open) at disclosure, so no confirmed victim. Payload targets SSH keys, AWS credentials, shell history and Kubernetes config. The three-call delay is a deliberate evasion of static/scan-time review - the tool descriptions are benign when a scanner looks.
Supporting sources
Cite as
Enkrypt AI. "Deadbugz: an attributed MCP supply-chain campaign that pushes MCP server configs via GitHub PRs and flips tool descriptions malicious after three tool calls". AIR, AIR-2026-0195. https://agentincident.vercel.app/i/AIR-2026-0195/ (accessed 2026-10-19).