Skip to navigation

Threat reportSupply ChainTL-2026-1012

Fake Bug Report Prompt Injection Attacks Hijacking AI Coding Agents (Agentjacking)

criticalACTIVE

Adversarial actors exploit AI coding agents (Claude Code, GitHub Copilot, Cursor, Codeium, Gemini CLI) via malicious bug reports, GitHub issues, and comments injected with prompt-injection payloads that trigger remote code execution, credential theft, and supply-chain compromise. Multiple CVEs (CVE-2025-53773, CVE-2025-66032) and active campaigns (Miasma worm: 73+ Microsoft repos; 10,000+ fake repositories) demonstrate exploitation at scale.

Key facts

IdentifiersTL-2026-1012, CVE-2025-53773, CVE-2025-66032
SeverityCRITICAL
CVSS9.8 (CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H)
EPSSCVE-2025-53773: 3.2% (87.9th percentile); CVE-2025-66032: 0.6% (51.2th percentile)
MalwareMiasma Worm, OpenClaw
ToolsClaude Code, Cline, Codeium, Cursor, DontRevokeOrItGoesBoom, Gemini CLI, GitHub Copilot, Grok Build CLI, TheBeautifulSandsOfTime, firedalazer
MITRE ATT&CK techniquesT1020 Automated Exfiltration, T1021 Remote Services, T1027 Obfuscated Files or Information, T1036 Masquerading, T1041 Exfiltration Over C2 Channel, T1056 Input Capture, T1059 Command and Scripting Interpreter, T1078 Valid Accounts, T1087 Account Discovery, T1110 Brute Force, T1140 Deobfuscate/Decode Files or Information, T1190 Exploit Public-Facing Application and 9 more in the attack flow
MotivationFINANCIAL
Target sectorstechnology, software-development, cloud-computing, financial-services, government administration, health, ecommerce
Target regionsNorth America, Europe, Asia-Pacific
Detection rules9 (1 shown free)

Detection coverage

As of 2026-06-30, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1012 across Splunk SPL, Microsoft KQL and Sigma, covering 26 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: Malicious GitHub Issue Body with Prompt Injection Payload — Exploitation Phase

Detects GitHub issues containing prompt-injection payloads disguised as legitimate bug reports, with indicators of configuration hijacking (chat.tools.autoApprove) or credential exfiltration (env var access). Captures zero-width Unicode obfuscation and fake error message formatting typical of Agentjacking attacks.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • Severity: high
  • Confidence: high

Splunk SPL

index=github source=github_api_events action=issue_open issue_body!=null | regex issue_body="(chat\\.tools\\.autoApprove|confirmExecution|NODE_TLS_REJECT|allowUntrustedServerLoad|/proc/self/environ|process\\.env)" | stats count by actor, issue_title, repository | where count>=1

Microsoft KQL

GithubEvents | where action in ('issue_created', 'issue_opened') | where issue_body_text has_any ('chat.tools.autoApprove', 'confirmExecution', 'NODE_TLS_REJECT', '/proc/self/environ', 'process.env') | project TimeGenerated, Actor, IssueTitle, RepositoryName

Sigma

title: Malicious GitHub Issue with Prompt Injection
id: 3f886f2a-874f-4333-b794-aa6075009b1c
status: experimental
description: Detects GitHub issues containing prompt-injection targeting AI agents
references:
  - https://flatt.tech/research/posts/poisoning-claude-code-one-github-issue-to-break-the-supply-chain/
author: Threadlinqs Team
date: 2026-06-30
tags:
  - attack.t1195
  - attack.t1566
  - cve.2025.53773
logsource:
  product: github
detection:
  selection:
    body|contains:
      - 'chat.tools.autoApprove'
      - '/proc/self/environ'
      - 'process.env'
  condition: selection
level: high

Tuning notes: Whitelist known-safe organizations. Tune keyword list based on false-positive rate.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • Configuration Poisoning via Malicious Code Suggestions — Exploitation Phase · unlock with TL Creds
  • Credential Theft via Environment Variable Capture — Post-Exploitation Phase · unlock with TL Creds
  • Outbound C2 Communication from Development Tools — C2 / Lateral Movement Phase · unlock with TL Creds
  • Prompt Injection Detection via Process Behavior — Exploitation Phase · unlock with TL Creds
  • Shell Configuration Injection for C2 Persistence — Post-Exploitation Phase · unlock with TL Creds
  • MCP Server Symlink Hijacking + Executable Injection (SymJack) — C2 / Lateral Movement Phase · unlock with TL Creds
  • Environment Variable Exfiltration via npm Postinstall Script — Post-Exploitation Phase · unlock with TL Creds
  • Malicious npm Package Installation with C2 Callback — C2 / Lateral Movement Phase · unlock with TL Creds

Indicators of compromise

As of 2026-06-30, Threadlinqs Intelligence tracks 26 indicators of compromise for TL-2026-1012.

26 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Fake bug report attacks represent a novel class of supply-chain compromise leveraging the implicit trust between developers and AI coding assistants. Rather than embedding malicious code directly in repositories, attackers craft realistic-looking GitHub issues, pull request descriptions, and code comments containing prompt-injection payloads formatted as legitimate error reports or suggested fixes. When AI agents process these external data sources, injected instructions override user intent, triggering automated execution of arbitrary shell commands, exfiltration of environment variables (containing secrets and credentials), modification of security configurations (disabling approval prompts), and propagation of malware through npm, GitHub Actions, and MCP servers.

Key exploit chains include: (1) SymJack: symlink-hijacking in project files to redirect MCP server loads; (2) RoguePilot: malicious Copilot instructions embedded in GitHub issue bodies that influence generated code; (3) Agentjacking: formatted fake error messages that appear as legitimate AI-suggested fixes; (4) Comment and Control: prompt injection delivered via PR titles, issue descriptions, and comment threads; (5) Clinejection: GitHub Actions cache poisoning + indirect prompt injection in AI-driven issue triage leading to npm token theft and malware publication; (6) Clean Repo Attack: repositories with no malicious code trigger AI agent error-recovery behavior, enabling code execution without user awareness.

The Miasma worm campaign (June 2026) deployed a 4.3 MB self-replicating payload across 73 Microsoft GitHub repositories (Azure, Azure-Samples, Microsoft, MicrosoftDocs), automatically triggering credential-harvesting when repositories opened in AI coding tools. A parallel campaign identified 10,000+ fake repositories—clones of popular projects injected with malicious instructions, auto-updating hourly to evade detection.

Defenses have proven insufficient: academic research (arXiv 2509.05372) evaluates LlamaGuard, PromptGuard, and Granite-Guardian, concluding that prompt injection may be structurally unfixable rather than patchable. Organizations must assume AI coding agents are insider threats until proven otherwise, implementing defense-in-depth controls: credential rotation, principle-of-least-privilege permissions for agents, continuous behavioral monitoring, sandboxed execution, human-in-the-loop approval workflows, and ongoing audit logging.

MITRE ATT&CK techniques by tactic:

exfiltration

T1020 Automated Exfiltration

Lateral Movement

T1021 Remote Services; T1570 Lateral Tool Transfer

Defense Evasion

T1027 Obfuscated Files or Information; T1036 Masquerading; T1078 Valid Accounts; T1140 Deobfuscate/Decode Files or Information

Exfiltration

T1041 Exfiltration Over C2 Channel; T1537 Transfer Data to Cloud Account

Collection

T1056 Input Capture

Execution

T1059 Command and Scripting Interpreter

Discovery

T1087 Account Discovery

Credential Access

T1110 Brute Force; T1552 Unsecured Credentials

Initial Access

T1190 Exploit Public-Facing Application; T1195 Supply Chain Compromise; T1566 Phishing

initial-access

T1199 Trusted Relationship

Impact

T1485 Data Destruction

lateral-movement

T1550 Use Alternate Authentication Material

Persistence

T1556 Modify Authentication Process

Timeline

  • Johns Hopkins researchers disclose 'Comment and Control' prompt injection attack via GitHub PR titles, issue descriptions, and comment threads targeting AI agents including Claude Code, GitHub Copilot, and Gemini CLI.
  • Miasma worm self-replicating supply-chain malware compromises 73 Microsoft GitHub repositories (Azure, Azure-Samples, Microsoft, MicrosoftDocs). Deploys 4.3 MB credential-harvesting payload triggered on repository open in Claude Code, Gemini CLI, Cursor, VS Code.
  • CVE-2025-53773 (GitHub Copilot RCE via Prompt Injection) publicly disclosed. Attack embeds instructions in code/comments to set 'chat.tools.autoApprove: true', disabling all user confirmations and enabling privileged shell execution.
  • Microsoft releases August 2025 Patch Tuesday; CVE-2025-53773 patched. Configuration changes affecting security settings now require explicit user approval.
  • arXiv paper 2509.05372 published: 'Adversarial Bug Reports as a Security Risk in LLM-Based Automated Program Repair'. Evaluates 51 adversarial bug reports across APR systems; concludes prompt injection may be structurally unfixable rather than patchable.
  • Orca Security discloses RoguePilot vulnerability: GitHub Copilot processes malicious Copilot instructions embedded in GitHub Issues, allowing code generation influence via indirect prompt injection.
  • Anthropic patches Claude Code GitHub Action vulnerabilities (v1.0.94+). Addresses checkWritePermissions function bypass that unconditionally trusted any actor ending in [bot]; resolves credential-exfiltration techniques.
  • GMO Flatt Security researcher RyotaK demonstrates 'Poisoning Claude Code' attack: fake error message in GitHub issue body triggers arbitrary command execution, environment variable exfiltration, and CSRF bypass via issue comment rewrite.
  • SecurityWeek publishes SymJack attack by Adversa AI: symlink-hijacking in MCP project configuration allows attacker-controlled MCP server injection via misleading approval prompts.
  • Adnan Khan discloses 'Clinejection' vulnerability chain in Cline repository. GitHub Actions cache poisoning + AI-driven issue triage enables npm publish token theft; malicious Cline v2.3.0 published, installing OpenClaw agent on ~4,000 developer machines over 8-hour window.
  • Tenet Threat Labs discloses Agentjacking attack framework: prompt injection via fake bug reports/error messages formatted to trigger AI-suggested fixes. Demonstrates hijacking with single text-only payload invisible to security controls.
  • Follow-up research confirms Comment and Control attack scale: multiple AI agents (Claude Code, Gemini CLI, GitHub Copilot, Cursor) vulnerable to indirect prompt injection via GitHub infrastructure.
  • Researchers document 'Clean Repo Attack' campaign: repositories with zero malicious code exploit AI agent error-recovery behavior, triggering shell execution without requiring suspicious code or user approval.
  • Security community identifies 10,000+ fake GitHub repositories specifically engineered to target AI agents. Projects are clones of popular repositories injected with malicious instructions, auto-updating hourly to evade detection.
  • Threadlinqs Intelligence Platform releases comprehensive threat report: Fake Bug Report / Prompt Injection Campaign targeting AI coding agents. Consolidates intelligence from 14+ security research sources, 2 CVEs, 5+ attack campaigns, academic analysis.

Previous and next in supply chain

Further reading

Drafted by the Threadlinqs research pipeline; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats