Skip to content

Security Alert | Open-Source Autonomous Pentesting Tool AntiFox Used in Real Intrusions, Pulls Live AWS Production Secret

Nov 13, 20261 min
TL;DRArctic Wolf found AntiFox, a fork of ARTEX, being used in real attacks: compromise an IIS server with a webshell, read credentials out of its config files, then use the stolen AWS identity to enumerate a dozen-plus services, landing a successful GetSecretValue call against a production secret. A separate intrusion read, copied, and deleted S3 objects. AWS says its services weren't compromised — the gap was long-lived keys sitting in customer application configs. Fixes: strip long-lived keys out of files like web.config, enforce IMDSv2, and alert on Secrets Manager access.

Look up & save

Loading dictionary…

🌏 中文版

What happened

On November 10, security firm Arctic Wolf published threat intelligence on AntiFox, an open-source autonomous pentesting platform hosted on Alibaba Cloud, showing it had moved from authorized testing into real intrusions. In one case, attackers compromised an internet-facing IIS web server, dropped an ASPX webshell, and ran OS commands to read database connection strings and credentials out of the application's config files. In another, attackers simply started with a valid AWS identity obtained by unknown means. From there, infrastructure linked to AntiFox fired off a large volume of cross-service AWS API calls using those identities — most got AccessDenied, but a handful succeeded. At least one intrusion progressed all the way to a successful GetSecretValue call against a production secret; another used a valid identity to read, copy, and delete S3 objects. The report was independently covered the same week by The Hacker News, BleepingComputer, and Security Affairs, and AWS responded on November 12 confirming its services weren't breached — the gap was customers leaving long-lived keys in application configuration.

Key facts

ItemValue
Incident typeAutonomous pentesting tool abused for real attacks (IIS webshell intrusion + AWS credential theft and lateral enumeration)
ScopeInternet-facing IIS applications with long-lived AWS credentials left in config files; at least one confirmed production secret leak and one incident of S3 objects being read and deleted
SeverityHigh
CVENone (Arctic Wolf did not identify the exact vulnerability used for code execution on the IIS side)
SourcesArctic Wolf Labs report, The Hacker News, BleepingComputer

Attack surface analysis

AntiFox is a fork of ARTEX 0.1.1 with a Chinese-language interface, built to let a planner/worker/main-agent multi-agent system autonomously run a full pentest — recon, attack-surface discovery, exploitation, evidence capture, and report delivery. The observed attack chain had two stages. The first was a conventional web compromise: once an internet-facing IIS server was breached, w3wp.exe (the IIS worker process) created an ASPX webshell and used that same process to run whoami and icacls to confirm permissions and write access, then used PowerShell to parse connection strings and app settings out of web.config, pulling a credential for an internal production database. Arctic Wolf didn't pin down the exact vulnerability used to get code execution, but a webshell created by w3wp.exe followed by commands run through the same process is a strong intrusion signal on its own.

The second stage was cloud identity abuse. Once in possession of a valid AWS identity — an IAM user, or in some cases a hijacked EC2 role — the attack infrastructure fired hundreds of cross-service API requests across multiple identities in roughly 30 minutes, touching IAM, EC2, S3, Lambda, Secrets Manager, KMS, RDS, ECS, EKS, Redshift, Athena, Glue, CloudWatch, CloudFormation, DynamoDB, and even Amazon Bedrock. Most calls came back AccessDenied, but the successful ones buried inside that same burst are what actually caused damage: at least one case progressed from enumerating Secrets Manager to a successful GetSecretValue call against a production secret, and another used a valid identity to run ListBuckets → ListObjects → GetObject on S3, followed by CopyObject and DeleteObject.

Mapped to the OWASP LLM Top 10, this sits closest to LLM05 Supply Chain Vulnerabilities — long-lived keys sitting in application config are the root cause that let the chain jump from a web compromise to the cloud — compounded by LLM06 Excessive Agency: an autonomous agent automates the judgment call of "given this credential, which services and APIs are worth probing," letting an attacker exhaustively test a dozen-plus services within half an hour. Even if most of those attempts fail, the few that succeed still do real damage. AWS's November 12 statement draws the line plainly: "AWS services were not compromised and operated as designed. The research describes threat actors using credentials taken from customers' compromised applications to access their AWS accounts." That echoes two incidents this site has already covered — Storm-0940's ARTEX-driven ConnectWise supply-chain attack and WorkOS's ARTEX sandbox-escape incident: the autonomous pentesting agent itself isn't the vulnerability — conventional appsec lapses (exposed services, long-lived keys in config, overly broad IAM permissions) are — but once an agent that can chain multi-step attacks autonomously and without fatigue runs into those lapses, the time between "finding a weakness" and "walking off with production data" compresses sharply.

AntiFox attack chain: from IIS webshell to AWS secret theft
AntiFox attack chain: from IIS webshell to AWS secret theftAttackers either compromise an IIS server or already hold a leaked AWS identity, then use an autonomous agent to exhaustively probe a dozen-plus cloud services; most requests are denied, but a few succeed in pulling a production secret or tampering with S3 objects. Breach internet-facing IIS w3wp.exe drops ASPX webshell Attacker holds AWS creds Read creds from web.config whoami / icacls confirm access Enumerate AWS services at scale Hundreds of calls in ~30 min Request succeeds? cross-service API sweep Most requests (AccessDenied) GetSecretValue succeeds production secret leaked S3 objects read, copied, deleted most a few succeed

Simplified from the Arctic Wolf Labs report.

Defenses

Immediate actions

  • Check config files on internet-facing web applications (IIS in particular) — web.config, deployment settings, debug/error logs — for long-lived AWS access keys, database passwords, or other plaintext credentials; treat any found as already leaked, revoke and rotate immediately
  • Check whether w3wp.exe has spawned child processes like cmd.exe, powershell.exe, whoami.exe, or icacls.exe, and whether writable web directories contain unexpected .aspx files
  • In CloudTrail/GuardDuty, alert on bursts of requests from unfamiliar hosting providers or ASNs using multiple IAM identities in a short window, especially a GetSecretValue call preceded by a run of access-denied enumeration

Long-term architecture

  • Enforce IMDSv2 on EC2 instance roles and restrict the metadata hop limit so role credentials can't be trivially lifted and abused
  • Apply least privilege to application identities — no single identity should be able to list and read across IAM, S3, Secrets Manager, and RDS at once
  • Use a cross-platform agent runtime governance tool like Netzilo to build a kill switch for which autonomous agent tools are allowed to make requests against production, instead of reconstructing it after the fact from CloudTrail

Impact

Arctic Wolf's report doesn't disclose how many organizations were affected, only confirming that among "several" intrusions, at least one successfully pulled a production secret and another had S3 objects read and deleted. AntiFox remains publicly downloadable on GitHub — now updated to version 2.0 with improved reporting — and its maintainers' documentation prohibits using it against real systems or without authorization, but enforces nothing technically. AWS's November 12 clarification that its services weren't breached places the responsibility on customer-side credential hygiene, which also means there's no official patch timeline for this class of incident — defense depends entirely on whether organizations clean the long-lived keys out of their own config files.

If your team has any internet-facing web application connected to AWS, assume right now that an autonomous pentesting agent may already be probing it exhaustively — what these tools lower isn't the technical bar for an attack, but the human time and effort it takes to try the same thing across many services and many identities at once.

Takeaway

The part worth remembering here isn't AntiFox itself — it's how AWS responded. The statement draws a clean line between "our services are fine" and "the customer left long-lived keys somewhere they shouldn't have." Autonomous agents speed up how fast an attack chain executes; they don't create the exposure. The accounts that got their secrets pulled already had that exposure — it just used to take manual, step-by-step probing to find it.

References