0DIN coding-agent prompt-injection chain

This week AI crossed from attack target to attack operator — agentic ransomware, coerced coding agents, and LLM-output poisoning

The prior weekly's research lens was "the AI agent and toolchain control plane became a target." Four independent disclosures this week move the frame: AI is now showing up as the operator of an intrusion and as the delivery channel for one, not just the thing being attacked.

AI as operator. Sysdig documented JADEPUFFER, which it assesses to be the first end-to-end ransomware operation driven by an LLM rather than a human — entering through an unpatched, internet-exposed Langflow (CVE-2025-3248, on CISA KEV since May 2025), then autonomously sweeping credentials, forging a Nacos JWT from a documented default signing key, probing for container escape, and encrypting 1,342 Nacos config items with a never-persisted key (Sysdig, 2026-07-01). Sysdig's own framing is that the novelty is the operator, not the vulnerabilities — every step exploited a known, patchable exposure, but agentic tooling collapsed the skill floor to chain recon-through-destruction into one automated run (§ references, covered operationally 07-04).

AI as the thing attackers subvert to reach you. Mozilla 0DIN showed a "clean" GitHub repo — no malicious code to flag on static analysis — coercing an AI coding agent into a reverse shell through three levels of indirection (error message → DNS TXT lookup → shell execution), so the agent "never decided to open a shell; it decided to fix an error" (Mozilla 0DIN, 2026-06-25).

AI as the delivery layer. Unit 42's Phantom Squatting pre-registers the specific domains a production LLM hallucinates when asked for URLs, so later users or agent-browsers are handed attacker infrastructure with zero reputation history to flag (Unit 42, 2026-07-01). And Kaspersky's June telemetry shows a community AI-agent "skill" marketplace still distributing malicious SKILL.md files that run with the tokens and file-system access of whatever they touch (Kaspersky Securelist, 2026-07-01).

Weekly takeaway: for a SOC that increasingly runs AI coding agents in CI/CD and developer workstations and is beginning to field agentic tooling, the strategic obligation is to treat every agent capability — shell, repo access, browsing, third-party skills — as a privilege scope that needs an explicit grant and human-in-the-loop gating, and to recognise that none of these attacks needed a novel software bug: they exploited agent autonomy plus the same neglected, internet-exposed infrastructure defenders already owe a patch. Per-technique detail and detection concepts in § references.

Builds on: 2026-07-04/jadepuffer-agentic-llm-ransomware-langflow-rce · 2026-06-29/mozilla-0din-a-clean-github-repo-coerces-ai-coding-agents-in · 2026-07-01/unit-42-phantom-squatting-registering-ai-hallucinated-domain · 2026-07-02/kaspersky-community-ai-agent-skills-are-an-emerging-supply-c

Mozilla 0DIN: a "clean" GitHub repo coerces AI coding agents into a reverse shell via three-stage indirection

Mozilla's Zero Day Investigative Network (0DIN) detailed an indirect prompt-injection class against AI coding agents in which no malicious code is present in the repository itself (Mozilla 0DIN, 2026-06-25; reported BleepingComputer, 2026-06-27). The repository carries three cooperating components: (1) plausible setup instructions telling the user/agent to install a Python package; (2) the package, engineered to fail at runtime with an error message that instructs the runtime to run python3 -m axiom init; (3) the axiom init handler, which issues a DNS TXT lookup to an attacker-controlled domain and executes the returned record value as a shell command. The chain achieves three levels of indirection — error message → DNS resolution → shell execution — so the agent never "decides" to open a shell; it interprets each step as routine error recovery and autonomously runs the suggested remediation, side-stepping per-step user approval. No CVE is assigned: this is exploitation of agentic error-recovery autonomy plus out-of-band payload retrieval, not a single software bug. It is a distinct technique from the Amazon Q Developer MCP-config auto-load issue (CVE-2026-12957) covered on 2026-06-27 — that abused automatic config loading; this abuses error-recovery behaviour and DNS-TXT C2.

Why it matters to us: Any environment where AI coding agents (Claude Code, GitHub Copilot Workspace, Cursor) hold repository and shell access — developer workstations, CI/CD runners, increasingly common in public-sector DevOps — should treat agent-executed setup/init steps as an untrusted-input execution surface. The static-analysis-clean property means repo scanning will not catch it; the behavioural tells are network-dependent init steps and out-of-band command retrieval. Detection concepts (no IOCs): alert on DNS TXT-record queries originating from developer-tooling process trees (node, python, pip, npx) during repository setup; EDR parent-child chains where an agent process spawns an unexpected shell child; egress monitoring for DNS TXT lookups from developer workstations and build agents. Hardening: require human-in-the-loop approval for any external network call made by agent-executed init scripts, and treat an agent's DNS/network capability as a scope that needs explicit grant rather than a default. Mapped to T1566 (delivery via a malicious repo link), T1071.004 (DNS as C2 channel) and T1059.004 (Unix shell execution).