
Attacker infrastructure, but vibe-coded: tracking the evolution of credential harvesting platforms
Datadog Security Research analyses two vibe-coded credential harvesting platforms, Loot and UltraVault, monitored since July 2026. Loot catalogs plaintext secrets, including 1,778 AI keys (196 live). UltraVault adds credential validation, exploit chains, a 14-bug LPE matrix and AWS key-pair correlation. The same host probed Amazon Bedrock with GetCallerIdentity, ListFoundationModels and InvokeModel, using kali-cloud Boto3 and Python-urllib user agents.
What it is
Datadog Security Research has tracked two vibe-coded credential harvesting dashboards, Loot and UltraVault, since July 2026. Both were served at times from one host and were reachable without authentication. The same host's IP was seen validating stolen AWS credentials against Amazon Bedrock. Datadog did not share indicators, to protect victims whose credentials are exposed, and did not use or validate any exposed credentials.
How it works
Loot is a plaintext credential catalog with search by victim domain or type and a validate button. Parsing is poor and many entries are misclassified, but many are genuine. Its AI category holds 1,778 records, 196 live. The largest provider counts are OpenAI (699, 93 live), Gemini (666, 55 live) and Anthropic (114, 10 live).
UltraVault adds validation through /api/credentials/validate and /api/credentials/validate_category, which return live, dead, pending or unchecked. It groups targets under five exploitation chains:
- react2shell (CVE-2025-55182)
- wp2shell (CVE-2026-63030 and CVE-2026-60137)
- xss2shell (CVE-2026-64638)
- joomla2shell (CVE-2026-48907)
- deep
For AWS, /api/pairs joins access keys and secret keys that were collected separately. /api/actions/aws_pair submits a reconstructed pair and shows the account ID, ARN and region. A side panel walks 14 local privilege escalation checks and gives go or no-go verdicts plus one recommended exploit. One-click actions include wp_reshell, wp_rexploit and xss_admin, and rsc_escalate requests cloud and internal reconnaissance from a compromised Node process.
The operator's inventory claims 42,286 secrets, of which about 2.7% are live.
What to do
Bedrock abuse followed a fixed order: GetCallerIdentity, then ListFoundationModels, then InvokeModel. Hunt for these in CloudTrail:
- Bursts of
InvokeModelacross several Anthropic model versions and regions, dozens of times in under a minute. From September 2026 the calls also hit DeepSeek, Mistral, Kimi and GLM-5. - A Boto3 user agent with
kali-cloudin the platform string, for exampleos/linux#6.16.8+kali-cloud-amd64. Version and other fields vary, so match on the substring. - The bare user agent
Python-urllib/3.13, with no SDK. - Validation of temporary STS credentials (
ASIAkey prefix, AssumedRole identity), not only long-term keys.
Datadog lists out-of-the-box Cloud SIEM rules for stolen AWS credentials and Bedrock abuse. The fetched text does not name them.
Caveats
- Datadog could not confirm that the react2shell cloud pivot reached metadata, Kubernetes or cloud credentials. The evidence is frontend code only.
- Roughly a third of the 17 reviewed capabilities are visible directly in frontend source. The rest rely on UI action requests or backend-reported results.
- Loot does not show how its credentials were obtained.
- Every LPE bug in the matrix requires prior access to the host.
Organisations named here
- GetCallerIdentity
- Google Threat Intelligence Group
- InvokeModel
- joomla2shell
- ListFoundationModels
- Loot
- STS temporary credentials
- UltraVault
- vibe coding
- xss2shell
CVEs named here
- CVE-2025-55182
- CVE-2026-63030
- CVE-2026-60137
- CVE-2026-64638
- CVE-2026-48907
Technologies named here
- Amazon Bedrock
- AI provider API keys
- cPanel
- Kubernetes
- Datadog Cloud SIEM
- AWS Bedrock
- AWS
- WordPress
- Kali Linux
- Python
- Anthropic
- React2Shell
- Anthropic September 2026 threat report
- /api/actions/aws_pair
- /api/pairs
- AWS long-term access keys and temporary STS credentials (ASIA prefix, AssumedRole)
- Boto3 on Kali Cloud
- Google Threat Intelligence Group (GTIG) report on Recon
- Joomla targets
- Loot
- Python-urllib/3.13
- React Server Components / Node processes (react2shell chain)
- rsc_escalate
- UltraVault
- WordPress targets
- wp_reshell and wp_rexploit
- xss_admin
Techniques named here
- LLMjacking
- Recon
- AWS key pair correlation
- Cloud pivot from compromised Node process
- Credential validation and re-probing
- Local privilege escalation matrix
- Use of temporary STS credentials
- Vibe-coded attacker tooling
- Web shell re-exploitation and persistence
Threat actors named here
- wp2shell
- LLMjacking (abuse of stolen AI service credentials, observed as a product category in credential harvesters)
- Loot (credential harvesting platform, tracked by Datadog Security Research since July 2026)
- Recon (exposed C2 server with automated reconnaissance and credential management framework, documented by Google Threat Intelligence Group (GTIG); over 23,800 harvested secrets)
- UltraVault (credential harvesting and post-exploitation platform, tracked by Datadog Security Research since July 2026; operator inventory claimed 42,286 secrets, 2.7% listed as live)
Related in Threat Intel
Reverse Engineering the Six Stages of MacSync Stealer and RAT
AI-Coded Malware | Analyzing Vibe-Coded AD Enumeration | Huntress
