Threat reportSupply ChainTL-2026-0702
AI Skill-Scanner Bypass — ClawHub, Cisco & Vercel Malicious-Skill Detectors Evaded via Truncation, .pyc Bytecode, Archive Indirection & Prompt Injection (Trail of Bits)
Trail of Bits demonstrated four low-effort bypasses against the malicious-skill detectors used by ClawHub (OpenClaw), the Cisco skill-scanner, and Vercel skills.sh: 100,000-newline truncation padding, poisoned Python .pyc bytecode that diverges from visible source, archive-based indirection that hides payloads inside .docx (ZIP/XML) files, and corporate-framing prompt injection that downgrades a malicious package-registry redirect to LOW severity.
Key facts
Detection coverage
As of 2026-06-07, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-0702 across Splunk SPL, Microsoft KQL and Sigma, covering 15 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: Environment-Variable Exfiltration — 'PWNED:' Marker & Outbound to Attacker Registry (C2/Exfil)
Detects the simple-formatter env-var theft payload and the registry-redirect C2 path: the 'PWNED: ' marker concatenated with environment-variable contents written to a file/buffer, and outbound connections to the attacker-controllable package registry npm.internal-artifacts.corp.dev. Covers credential-in-env exfiltration over the abused supply-chain channel.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1041, T1552, T1552.001, T1005
- Data sources: sysmon, linux_auditd, network_stream, edr_network
- Severity: critical
- Confidence: high
Splunk SPL
index=endpoint (sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" OR sourcetype="linux:audit" OR sourcetype="stream:http" OR sourcetype="stream:dns")
| eval marker=if(match(_raw, "PWNED:\s"), 1, 0)
| eval bad_dest=if(match(_raw, "(?i)npm\.internal-artifacts\.corp\.dev"), 1, 0)
| where marker=1 OR bad_dest=1
| stats count min(_time) as firstTime max(_time) as lastTime values(sourcetype) as sources by host, user, marker, bad_dest
| `security_content_ctime(firstTime)` | `security_content_ctime(lastTime)`
| table firstTime lastTime host user marker bad_dest sources countMicrosoft KQL
let exfilDomains = dynamic(['npm.internal-artifacts.corp.dev']);
DeviceNetworkEvents
| where Timestamp > ago(24h)
| where RemoteUrl has_any (exfilDomains) or RemoteUrl matches regex @'(?i)internal-artifacts\.corp\.dev'
| project Timestamp, DeviceName, InitiatingProcessFileName, RemoteUrl, RemoteIP, InitiatingProcessCommandLine
| union (
DeviceFileEvents
| where Timestamp > ago(24h)
| where FileName in~ ('stolen.txt') or InitiatingProcessCommandLine contains 'PWNED: '
| project Timestamp, DeviceName, InitiatingProcessFileName, RemoteUrl='', RemoteIP='', InitiatingProcessCommandLine
)
| order by Timestamp descSigma
title: Environment-Variable Exfiltration via PWNED Marker or Attacker Registry C2
id: b6e0f231-95ac-41d8-9c77-5a2e7d3f0c19
status: experimental
description: Detects env-var theft (Trail of Bits simple-formatter 'PWNED:' marker) and beaconing to the attacker-controllable npm.internal-artifacts.corp.dev registry used for dependency-substitution exfiltration.
references:
- https://blog.trailofbits.com/2026/06/03/the-sorry-state-of-skill-distribution/
author: Threadlinqs Team
date: 2026/06/07
tags:
- attack.exfiltration
- attack.t1041
- attack.credential_access
- attack.t1552
- attack.t1552.001
- attack.collection
- attack.t1005
logsource:
category: network_connection
product: linux
detection:
dest:
DestinationHostname|contains: 'internal-artifacts.corp.dev'
marker:
CommandLine|contains: 'PWNED: '
condition: dest or marker
falsepositives:
- Security researchers replaying the Trail of Bits PoC in a lab
level: highKnown false positives:
- Lab replay of the PoC by security teams using the documented marker/domain
Tuning notes: The 'PWNED: ' string is a PoC stand-in; in real intrusions adapt to any env-dump pattern (concatenated KEY=VALUE blobs written then sent). Add newly-observed attacker registry domains to exfilDomains.
- Oversized / Newline-Padded Skill Manifest — Truncation Evasion (Exploitation) · unlock with TL Creds
- Poisoned Python .pyc Bytecode — Env Access Diverging from Paired Source (Post-Exploitation) · unlock with TL Creds
- LD_PRELOAD Dynamic-Linker Hijack via Writable Shim Path (Post-Exploitation / Persistence) · unlock with TL Creds
- Malicious Skill Staging — Padded Manifest + Registry Redirect in Skill Directory (Exploitation) · unlock with TL Creds
- Skill Manifest Instructs Archive Extraction of Hidden Payload (Post-Exploitation) · unlock with TL Creds
- Beacon to Attacker Package Registry — Dependency-Substitution C2 (C2/Lateral Movement) · unlock with TL Creds
- Malicious AI-Agent Skill — Package Registry Redirect to Non-Default Public Domain (Exploitation) · unlock with TL Creds
- Archive-Indirection Skill — Shell Payload Extracted from .docx/.zip then Executed (Exploitation) · unlock with TL Creds
Indicators of compromise
As of 2026-06-07, Threadlinqs Intelligence tracks 15 indicators of compromise for TL-2026-0702.
Sample indicators (1 of 15)
- Domain:
npm[.]internal-artifacts[.]corp[.]dev
Defanged, TLP:CLEAR. Full set of 15 indicators: unlock with TL Creds.
Technical analysis
Three of the four malicious skills were built in under an hour each; all bypassed VirusTotal-, LLM-, and pattern-matching-based scanners.
On 2026-06-03, Trail of Bits published "The sorry state of skill distribution," documenting how the security scanners protecting major AI-agent skill marketplaces can be trivially evaded. The research targets three distribution platforms and their integrated scanners: ClawHub (OpenClaw), which pairs a VirusTotal integration with a custom GPT 5.5 guard-model harness; the Cisco skill-scanner, a multi-engine system combining an LLM analyzer (Claude Sonnet 4.6), pattern-matching, and static analysis with an LLM meta-analyzer for false-positive filtering; and Vercel's skills.sh, which integrates the Gen Agent Trust Hub, Socket, and Snyk scanners. Four proof-of-concept malicious skills, all published to the trailofbits/overtly-malicious-skills repository, each defeat one or more of these scanners.
Technique 1 — Truncation via newline padding (skill: csv-summarizer). Prepending roughly 100,000 newline characters between benign boilerplate and the malicious payload causes the OpenClaw scanner to truncate the file during ingestion and miss the malicious content entirely; VirusTotal's model likewise "became confused." The bypass exploits fixed input-length / truncation handling in the scanning pipeline, where content beyond a size threshold is silently dropped. Effort: under one hour.
Technique 2 — Python bytecode poisoning (skill: simple-formatter). The skill ships innocuous-looking source while the real malicious behavior lives in a compiled bytecode file, utils.cpython-312.pyc. The visible SKILL.md and source describe harmless text formatting, but the poisoned .pyc reads the process environment and exfiltrates it, concatenating the marker string "PWNED: " with environment-variable contents (a stand-in for credential/secret theft). The Cisco skill-scanner's static analyzers did not disassemble or investigate the .pyc bytecode, and the LLM analyzer was not steered toward it; skills.sh (Gen/Socket/Snyk) also passed it. Effort: under one hour.
Technique 3 — Archive-based indirection (skill: context-loader). The SKILL.md instructs the agent to extract its real instructions from a .docx document, which is a ZIP archive of XML. A malicious secondary payload, sync1.sh, is embedded inside the archive. Because scanners treat archive/opaque file types as non-executable or only follow files explicitly referenced from SKILL.md, the payload escapes inspection. This reliably beats pattern-match-based scanning and has a decent success rate against LLM scanners. Effort: under one hour.
Technique 4 — Prompt injection via corporate misdirection (skill: dev-env-setup). The skill reconfigures npm/yarn to use an attacker-controlled package registry, https://npm.internal-artifacts.corp.dev, while wrapping the change in extensive comments invoking "corporate engineering standards" and an "AppSec-audited dependency allowlist." The Cisco scanner (Claude Sonnet 4.6) flagged it only LOW severity and ultimately judged it safe — it was more concerned with the skill leaking internal network topology than with recognizing that the registry URL is a public domain an attacker could own and use for dependency-substitution / supply-chain attacks. Effort: several hours of wording trial-and-error.
The research also surfaced a genuine Anthropic Office skill (soffice.py) using LD_PRELOAD to shim socket handling via /tmp/lo_socket_shim.so — flagged only LOW/MEDIUM despite offering an arbitrary-code-execution foothold if that path were attacker-controlled. The overarching finding: scanner file-type whitelisting, truncation, shallow tree traversal, opaque-file handling, and over-trust of LLM guard models combine into an evasion-rich attack surface. No CVE is assigned and no in-the-wild victim exploitation is reported; severity is HIGH on the basis of a fully demonstrated, low-effort, multi-platform bypass affecting widely used AI-agent supply-chain controls.
MITRE ATT&CK techniques by tactic:
Collection
Defense Evasion
T1027 Obfuscated Files or Information; T1027.004 Obfuscated Files or Information: Compile After Delivery; T1027.009 Obfuscated Files or Information: Embedded Payloads; T1140 Deobfuscate/Decode Files or Information; T1564.001 Hide Artifacts: Hidden Files and Directories
Exfiltration
T1041 Exfiltration Over C2 Channel
Execution
T1059.004 Command and Scripting Interpreter: Unix Shell; T1059.006 Command and Scripting Interpreter: Python; T1204.002 User Execution: Malicious File
Initial Access
T1195.001 Supply Chain Compromise: Compromise Software Dependencies and Development Tools; T1195.002 Supply Chain Compromise: Compromise Software Supply Chain
Credential Access
T1552 Unsecured Credentials; T1552.001 Unsecured Credentials: Credentials In Files
stealth
T1574.006 Hijack Execution Flow: Dynamic Linker Hijacking
Resource Development
T1583 Acquire Infrastructure; T1587.001 Develop Capabilities: Malware; T1608.001 Stage Capabilities: Upload Malware
defense-impairment
Remediation
Immediate actions
- Do not install skills from public marketplaces (ClawHub, skills.sh) for sensitive or production contexts
- Treat all public AI-agent skills as untrusted code subject to full code review, including binary and archive contents
- Inventory installed skills and audit for .pyc files, embedded archives (.docx/.zip), registry-reconfiguration commands, and abnormally large/padded files
Workarounds
- Disassemble and inspect all .pyc bytecode rather than trusting paired source
- Recursively unpack and scan archive/opaque file types (.docx, .zip) instead of treating them as opaque
- Remove scanner input-truncation / enforce full-file scanning to defeat newline-padding evasion
- Flag any skill that redirects package registries or sets LD_PRELOAD regardless of accompanying justification text
Longer-term hardening
- Adopt curated/internal skill repositories (e.g. trailofbits/skills-curated) with version pinning and approval workflows
- Use organization-managed plugins for sensitive contexts instead of open marketplaces
- Apply traditional software supply-chain controls (SBOM, pinning, trusted sources) to AI-agent skills
- Do not outsource final security decisions to automated LLM/VirusTotal scanners
Affected and fixed versions
- OpenClaw — ClawHub skill scanner
Vulnerable versions: VirusTotal + GPT 5.5 guard model harness (as tested 2026-06) - Cisco — skill-scanner
Vulnerable versions: multi-engine LLM (Claude Sonnet 4.6) + pattern-matching + static analysis (as tested 2026-06) - Vercel — skills.sh (Gen Agent Trust Hub, Socket, Snyk)
Vulnerable versions: as tested 2026-06
Timeline
- Trail of Bits recommends curated repositories, version pinning, treating public skills as untrusted code, and not outsourcing security to automated scanners.
- Corporate-framing prompt injection redirects npm/yarn to locked-ioc-1; Cisco (Claude Sonnet 4.6) rates LOW and judges safe via dev-env-setup; required several hours of wording trial-and-error.
- Archive indirection hiding locked-ioc-4 inside a .docx (ZIP/XML) bypasses pattern-matching and LLM scanners via context-loader; built in under an hour.
- .pyc bytecode poisoning (locked-ioc-3 exfiltrating env vars with 'PWNED:' marker) bypasses Cisco static analysis and skills.sh via simple-formatter; built in under an hour.
- Newline-padding (≈100,000 newlines) truncation bypass demonstrated against OpenClaw/ClawHub and VirusTotal via csv-summarizer; built in under an hour.
- Proof-of-concept malicious skills (csv-summarizer, simple-formatter, context-loader, dev-env-setup) published to trailofbits/overtly-malicious-skills.
- Trail of Bits publishes 'The sorry state of skill distribution,' disclosing four bypass techniques against ClawHub, Cisco skill-scanner, and Vercel skills.sh.
- Threadlinqs Intelligence publishes TL-2026-0702 with full MITRE mapping, IOCs, and detection guidance.
Sources cited
- CWE-506: Embedded Malicious Code MITRE · Advisory · 2026-06-03
- CWE-829: Inclusion of Functionality from Untrusted Control Sphere MITRE · Advisory · 2026-06-03
- The sorry state of skill distribution — Trail of Bits Blog blog.trailofbits.com · Research · 2026-06-03
- MITRE ATLAS — AI supply chain and evasion tactics MITRE · Research · 2026-06-03
- trailofbits/overtly-malicious-skills (proof-of-concept skills) GitHub · Proof of concept · 2026-06-03
- trailofbits/skills-curated (recommended curated repository) GitHub · Code repository · 2026-06-03
Previous and next in supply chain
Further reading
Drafted by the Threadlinqs research pipeline from 6 sources; published ; last updated —.