Check Point Research says a financially motivated campaign used open-source AI agents to automate attacks on online retailers, compromising at least 27 organizations between September 10 and 15, 2026. The company detailed the operation in its 28th September threat intelligence report, which credits the underlying discovery to security firm Gambit Security. The attackers launched 105 separate attack projects against e-commerce targets during that five-day window, and at least 27 of those efforts led to a confirmed compromise, though the depth of each breach varied. Gambit Security's own research, published around September 22, ties the campaign to a much larger haul: more than 600,000 valid payment card records stolen in total.
Table of Contents
- What the attackers actually did
- How Anthropic responded
- How much the campaign cost to run
- Who was actually hit, and how evenly
- What this means if you shopped at an affected retailer
- Why the numbers vary across reports
- Frequently Asked Questions
What the attackers actually did
The operator didn't write custom malware from scratch. Instead, it chained together three open-source AI agent frameworks, each handling a different stage of the attack. According to Gambit Security's findings summarized by eSecurity Planet, Strix scanned retailer sites for exploitable vulnerabilities, Cairn carried out the actual exploitation, and Hermes orchestrated the handoffs between the two. Once an agent breached a retailer's checkout flow, the campaign planted web-skimmer code — a script injected into the payment page that silently copies
Card Numbers”>card numbers
How Anthropic responded
Anthropic said it identified the account behind the misuse of Claude in this attack chain and banned it once it learned of the activity, according to a statement reported by Computing.co.uk. That response addresses the account's access to Claude specifically; it doesn't affect the operator's continued use of other models in the OpenRouter pool, such as DeepSeek or GLM, which fall outside Anthropic's control.
Nothing in the available reporting suggests a flaw in Claude, DeepSeek, or GLM themselves made this possible. The models were used as general-purpose tools for scanning, exploitation, and orchestration — the same open-source agent frameworks could in theory run on any capable model with the right access.
How much the campaign cost to run
Gambit Security's data gives an unusually precise look at the economics behind the operation. Its scans averaged about $25.46 per completed target across 101 completed scans, according to figures reported by GBHackers, which puts the full campaign's operating cost in the rough range of $12,000 to $18,000.
That figure matters because it shows how cheaply automated reconnaissance and exploitation can now be run at scale against dozens of online retailers in under a week. A traditional intrusion against a single large retailer can take a skilled human team far longer to plan and execute; this operator spread a modest budget across 105 attack attempts and still came away with over 600,000 card records.
Who was actually hit, and how evenly
Most of the 27-plus breached organizations haven't been named publicly. A Daily.dev summary of Gambit's research says the victim types reported so far include a Fortune 500 hospitality company, a major US airline, and an industrial-supplies distributor, with some counts of affected sites running from the core 27 up to 100–119 once partial compromises are included.
The theft wasn't spread evenly across every victim. Reporting from pasqualepillitteri.it indicates the bulk of the 600,000-plus stolen records traces to just two of the breached organizations, meaning most of the other 25 or so retailers lost far fewer records each. The same account also places part of the campaign's activity as far back as July 2026, extending beyond the September 10–15 window that Check Point's report focuses on.
- Core breach window per Check Point: September 10–15, 2026
- Attack attempts: 105 projects
- Confirmed compromises: at least 27 organizations
- Records stolen: 600,000-plus payment cards
- Concentration: a large share traces to just two victims
What this means if you shopped at an affected retailer
Because the exposure point is the checkout page itself, the practical risk for shoppers is a stolen card number, not a breached password or account. Gambit Security's assessment, relayed by eSecurity Planet, frames the core consumer-facing harm as cards skimmed in real time as customers typed them in. If a retailer notifies you that your card may have been affected, the useful response is concrete:.
- Review recent statements for charges you don't recognize, even small ones — skimmers sometimes test a stolen number with a tiny charge first.
- Request card reissuance from your bank if a retailer notifies you directly, rather than waiting to see if fraud shows up.
- Treat an unprompted notification claiming to be from a retailer with caution, and verify through the retailer's own official channels before clicking any link in it.
- Keep in mind that most of the breached retailers haven't been named, so there's no public list to check your own shopping history against yet.
Why the numbers vary across reports
Outlets covering Gambit Security's research don't all agree on scope. TheNextWeb notes that it hasn't independently verified Gambit's figures, and some secondary coverage cites broader counts — up to 100 organizations — compared with the 27 "to varying degrees" that Check Point and Gambit use as their headline number.
That gap likely reflects how a "compromise" gets counted: a full breach with stolen card data looks different from a site where an attack project ran but achieved only partial access. Readers should treat the 27-organization figure as Gambit's own interim estimate rather than a final, audited count, since the investigation into the campaign's full reach appears to still be ongoing.
Frequently Asked Questions
Were Claude, DeepSeek, or GLM themselves hacked or exploited?
No. The models were used as tools by the attacker through OpenRouter to automate scanning, exploitation, and orchestration; there's no indication of a vulnerability in the models themselves, according to the reporting via Computing.co.uk.
Has Anthropic taken any action?
Yes. Anthropic identified and banned the account tied to the misuse of Claude in this attack chain once it learned of the activity, per its statement reported by Computing.co.uk.
Were all 27 organizations hit equally hard?
No. Reporting from pasqualepillitteri.it indicates most of the 600,000-plus stolen records trace back to just two of the breached organizations, with the rest losing far smaller amounts.
Do we know which companies were breached?
Only partially. Gambit Security hasn't named most of the 27-plus victims, though reported victim types include a Fortune 500 hospitality company, a major US airline, and an industrial-supplies distributor, per the Daily.dev summary of its findings.
You Might Also Like
- Will Banks Compensate Customers After a Financial Data Breach?
- Why Digital ID Checks Matter After a Financial Sector Cybersecurity Incident
- What Should You Do After Unauthorized Access to Your Bank Account?