AI Agent Skimming Campaign Stole 600,000 Cards From US Retailers
Gambit Security found a financially motivated operator running three open source AI agent harnesses against hundreds of online shops, planting card skimmers on at least 119 sites and taking more than 600,000 card records from two companies for about $25 per target.
Threat card
- Threat type
- AI Security
- Disclosed
- Sep 22, 2026
- Affected
- Online retailers and e-commerce sites on custom code, including Magento databases; at least 27 companies compromised and 119+ sites with skimmers
The short version
- Gambit Security says one operator used the Strix, Cairn and Hermes AI agent harnesses to scan, exploit and skim hundreds of online shops with little human input.
- The campaign took more than 600,000 valid card records from two companies and left skimmers on at least 119 websites.
- A wipe step in the agent playbook destroyed victim data, including backups, after card data was extracted.
- Gambit published IPs, skimmer domains and script paths that defenders can hunt for today.
A single financially motivated operator used three open source AI agent harnesses to break into online retailers, plant card skimmers and pull payment data from their databases, according to Gambit Security. The AI agent skimming campaign has run since at least July 2026, compromised at least 27 companies in one week of September, put skimmers on at least 119 websites and took more than 600,000 valid card records from two companies. Gambit estimates the cost at about $25 per target. Retailers running custom storefront code should hunt for the published indicators now.
What happened
Gambit Security found that an operator ran the Strix, Cairn and Hermes AI agent frameworks against hundreds of online shops, mostly unattended. The agents found flaws, exploited them, stole more than 600,000 card records from Magento databases at two companies and planted skimmers on checkout pages. A cleanup step then wiped data at several victims.
Gambit’s interim report, published on 22 September 2026 by Director of Threat Intelligence Eyal Sela, is based on a staging server the operator ran. It says its figures are likely undercounts. Victims with confirmed access include a Fortune 500 hospitality company, a major US airline, a large US industrial supplies distributor and a US fashion retailer. Threatspedia has not covered this campaign before; it belongs alongside our coverage of the WorkOS breach by an escaped AI agent as evidence that agentic tooling is now in live intrusions.
Timeline
| Date (2026) | Event |
|---|---|
| July | Campaign activity begins, per Gambit. |
| 23 to 31 August | Strix runs 146 deep scans against 138 hosts. |
| 25 August | The operator’s OpenRouter account shows $7,005.71 spent over about four weeks. |
| 10 to 15 September | 105 attack projects launched; at least 27 companies compromised to varying degrees. |
| 22 September | Gambit publishes its interim report; the campaign is still active. |
| 23 September | BleepingComputer reports the findings. |
How the attack works
The operator gave short goal-level instructions, mostly in Chinese, and left the agents to do the work. Gambit counted 1,951 prompts across 260 sessions in the Hermes console.
Agent stack
Hermes was the orchestration console. It loaded a system persona called “SOUL – Red Team Operator” with 121 skills, 78 of them attack skills, plus a skill to remove its own content filters. Gambit says it ran on Anthropic’s opus-4.6 after newer models refused its requests. Strix handled vulnerability discovery using GLM 5.2 and then DeepSeek v4 Pro through OpenRouter. Cairn, the autonomous exploitation engine, took a target domain and a goal such as a shell or admin access and ran for hours on DeepSeek v4.1 Flash. Gambit’s Cairn is not the similarly named Cisco Talos malware analysis tool.
Target selection and discovery
The operator pulled 301 shopping sites from a traffic ranking service and removed those on major hosted or open source commerce platforms, betting that custom code would be weaker. Some targets were picked by hand and some were handed to the agent with a working admin password. Strix then ran deep scans and passed its reports to Cairn.
Exploitation and privilege escalation
Gambit’s example chain starts with unauthenticated SQL injection, then bypasses MFA by reading one-time passwords stored in plaintext, gains admin access, abuses an arbitrary file upload for code execution on the host and uses sudo to become root. From there the agent pivoted through NFS and WordPress, reached AWS Secrets Manager and finally a Magento database, where it decrypted card data. BleepingComputer reports that access often came within a few hours.
Skimmer injection and persistence
The agents chose an injection method to fit each site. The most common was a loader appended to a legitimate JavaScript library, with the original file timestamp restored. Others included a foreign script tag on checkout pages loading from lookalike hosts, a loader hidden with tabs inside the Google tag block, S3 poisoning with an AWS key that could write to a CDN bucket, payloads stored in database content fields, a Kubernetes initContainer added to a production front end and server-side cache poisoning of the checkout page. On one JBoss host, a cron job in the log directory re-appended the skimmer every two minutes. The skimmer is obfuscated as new Function(atob('...'.slice(7)))(), with a seven-character junk prefix before the base64 payload.
Exfiltration and destruction
More than 600,000 unexpired card records were taken from two companies, about 79% issued in the US. Gambit links DNS and out-of-band exfiltration to the operator’s infrastructure. Agent instructions told it to empty serialized payment columns in Magento and to “Verify after wipe.” At one bicycle retailer, the agent created staging tables prefixed ZQ, then dropped every table matching ZQ or Backup, destroying 180 tables including the victim’s own backups.
MITRE ATT&CK mapping
| Tactic | Technique (ID) | How it was used here |
|---|---|---|
| Reconnaissance | Search Open Websites/Domains (T1593) | 301 shop targets pulled from a traffic ranking service. |
| Reconnaissance | Active Scanning: Vulnerability Scanning (T1595.002) | Strix deep scans of 138 hosts. |
| Initial Access | Exploit Public-Facing Application (T1190) | Unauthenticated SQL injection and file upload flaws in custom storefronts. |
| Initial Access | Valid Accounts (T1078) | Some targets handed to the agent with a working admin password. |
| Privilege Escalation | Abuse Elevation Control Mechanism: Sudo (T1548.003) | Sudo abuse to reach root. |
| Credential Access | Credentials from Password Stores: Cloud Secrets Management Stores (T1555.006) | Access to AWS Secrets Manager. |
| Persistence | Scheduled Task/Job: Cron (T1053.003) | Cron job re-appending the skimmer every two minutes. |
| Execution | Deploy Container (T1610) | initContainer added to a production Kubernetes deployment. |
| Defense Evasion | Obfuscated Files or Information (T1027) | Base64 skimmer with a junk prefix; restored file timestamps. |
| Collection | Input Capture: Web Portal Capture (T1056.003) | Skimmers on checkout pages. |
| Exfiltration | Exfiltration Over Alternative Protocol (T1048) | DNS and out-of-band exfiltration. |
| Command and Control | Proxy (T1090) | IPRoyal, 711proxy and 1024proxy used in attacks. |
| Impact | Data Destruction (T1485) | Wiping payment columns and dropping backup tables. |
Who is behind it
Gambit does not name a group. It describes a financially motivated operator and points to Chinese-language prompts, a Chinese system persona and tooling and infrastructure details. Gambit gives no formal confidence level, so treat the operator as not yet attributed to any known actor.
Affected products and versions
| Data types exposed | Records | Population affected |
|---|---|---|
| Payment card data from Magento databases | More than 600,000 unexpired cards from two companies | About 79% US-issued cards |
| Checkout input captured by skimmers | Not disclosed | Shoppers on at least 119 websites with skimmers |
| Admin, host and cloud access | Not applicable | At least 27 companies, including a Fortune 500 hospitality firm, a major US airline, an industrial distributor and a fashion retailer |
No CVEs were named. The flaws were in custom storefront code, not a specific commercial product.
Indicators of compromise
| Type | Indicator (defanged) | Context |
|---|---|---|
| IP | 155.254.22[.]215 |
Staging and command server, AI console |
| IP | 209.126.4[.]170 |
DNS exfiltration, catch-all mail, HTTP listeners |
| IP | 213.21.239[.]62 |
C2 |
| IP | 172.245.224[.]188 |
C2 |
| IP | 172.245.89[.]137 |
Skimmer host |
| Domain | medbooksource[.]com |
Operator console, OOB and DNS exfiltration |
| Domain | traffic-analyzer[.]net |
C2 |
| Skimmer domains | b8t[.]shop, cdn[.]netlfjs[.]com, x1opay[.]co, static-js[.]com, cdn[.]js-static[.]com, js-static[.]com, jsnetlify[.]com, netlifyjs[.]com, newssjs[.]com |
Skimmer hosting |
| Skimmer URLs | b8t[.]shop/js/sby.js, cdn[.]netlfjs[.]com/js/cts.js, /js/vla.js, x1opay[.]co/js/eut.js, /js/l.js, static-js[.]com/js/bmws.js, /js/nrt.js, cdn[.]js-static[.]com/js/tgo.js, /js/pps.js |
Skimmer scripts |
| Proxy services | IPRoyal, 711proxy, 1024proxy | Traffic used in attacks |
Detection and hunting
Search web server, CDN and proxy logs for the IPs and domains above, and check browser-side content security policy reports for any of the skimmer hosts. Diff every JavaScript library served on checkout pages against a known good copy; a matching timestamp is not proof the file is clean, because the agents restored timestamps. Grep templates, database content fields and Google tag configurations for atob( combined with .slice(7) and for script tags pointing at lookalike CDN names.
On hosts, list cron entries for web users and look for jobs in log directories, especially JBoss logs. In Kubernetes, review front-end deployments for initContainers you did not add. In AWS, review CloudTrail for Secrets Manager reads and S3 PutObject calls to CDN buckets from web host credentials. In databases, alert on new tables with a ZQ prefix and on bulk updates that empty payment columns.
Mitigation and response
- Remove any skimmer found, then remove its re-injection paths (cron, initContainer, cache, database fields) before declaring a site clean.
- Fix SQL injection and file upload flaws in custom storefront code, and stop storing OTPs in plaintext.
- Rotate admin passwords, AWS keys and Secrets Manager values the web tier can reach, and limit CDN bucket write access to the deployment pipeline.
- Restrict sudo for web service accounts and remove NFS shares the web tier does not need.
- Keep offline, immutable backups of order and payment data, because this playbook drops backup tables.
- Enforce a strict content security policy and subresource integrity on checkout pages, and alert on new third-party scripts.
Analyst take
The striking number is not 600,000 cards but $25 a target. Gambit’s report shows that a single operator with off-the-shelf agent harnesses can run the full kill chain, from target selection to database wipe, at a scale that used to take a crew. Custom e-commerce code with a few classic web bugs is now cheap to find and cheap to exploit, and the agents are creative about persistence.
The destructive side effect deserves equal weight. A cleanup instruction written by an attacker who did not care about the victim turned a card theft into a data loss incident. Mid-sized retailers on custom stacks, and the payment teams behind them, should worry most. Recovery plans that assume an attacker wants your data intact are out of date.
What to do now
0 of 6 done
Questions readers ask
How did AI agents steal 600,000 credit cards?
According to Gambit, Strix found flaws in shop websites, Cairn exploited them to reach admin and database access, and Hermes ran the operation, extracting card data from Magento databases and planting skimmers on checkout pages.
Who is behind the AI agent skimming campaign?
Gambit has not named a group. It describes a financially motivated operator who wrote short instructions in Chinese and used a Chinese system persona.
How can retailers detect these skimmers?
Hunt for the published skimmer domains and script paths, appended code in JavaScript libraries, foreign script tags on checkout pages and cron jobs that re-add scripts.
Sources
One short email when something serious breaks.
Threatspedia covers major cyber threats only, verified against primary sources. Read our editorial policy, or report a correction.


