AI Tools Are Generating Massive Attacks for Little Cost

Summary

Dark Reading reports that Gambit Security discovered a five-day September campaign in which an unnamed Chinese-speaking attacker targeted 105 e-commerce sites and compromised at least 27 companies. Reported outcomes included theft of at least 600,000 payment-card records from two companies, credit-card skimmer installation on five websites, and varying levels of illicit access elsewhere. At least one compromised company was described as a Fortune 500 organization. These figures come from a single supplied secondary report and are not independently corroborated in the evidence.^[0]

The attacker reportedly combined the Hermes agentic AI system with the Strix and Cairn penetration-testing frameworks, using both open-weight and public AI models. The reported average cost was $25.46 per completed target. The evidence establishes a common automated toolset and financially oriented outcomes, but does not disclose the initial-access techniques, specific vulnerabilities, victim software versions, or actionable attacker infrastructure. Attribution is limited to the description of a Chinese-speaking attacker; no named group is identified.^[0]

A separate Silverfort laboratory exercise illustrates the defensive implications without establishing an additional real-world compromise: an orchestrating agent using publicly available Claude Opus models reportedly reached full domain control from an initial foothold in eight minutes for about $12. After authentication and access controls were strengthened, the same agent reportedly encountered 21 dead ends and abandoned the attempt. The report's central defensive conclusion is that inexpensive automation increases pressure to harden exposed systems, close identity weaknesses, and prioritize continuity of essential business processes.^[0]

Attack chain

1. Automated targeting: Across a five-day September window, the attacker used three AI harnesses—Hermes, Strix, and Cairn—and a mixture of open-weight and public models to target 105 e-commerce sites. The report does not describe reconnaissance procedures or target-selection criteria.^[0]
2. Compromise and illicit access: The campaign reportedly compromised at least 27 companies, including at least one Fortune 500 organization. The evidence does not specify the entry vulnerabilities, access mechanisms, or privilege levels obtained at individual victims.^[0]
3. Payment-data theft and skimmer deployment: Following compromise, reported outcomes included theft of at least 600,000 payment-card records from two companies and installation of credit-card skimmers on five websites. These are separate reported outcomes; the evidence does not establish their overlap or the exact ordering of theft and skimmer installation.^[0]

Disclosure timeline

Key takeaways

• The campaign's reported significance is the combination of scale, payment-data theft, and low operating cost: 105 sites targeted, at least 27 companies compromised, and an average of $25.46 per completed target. The evidence does not provide a cost breakdown.^[0]
• Keep the observed retail campaign separate from the Silverfort experiment. The $12, eight-minute domain takeover was a controlled laboratory result, not a measured outcome of the campaign against e-commerce companies.^[0]
• The strongest concrete defensive result is the laboratory agent's failure after access controls were improved. This supports attention to existing identity weaknesses rather than treating AI-enabled attacks as requiring an entirely new class of vulnerability.^[0]
• The supplied evidence supports campaign-level awareness and control prioritization, but not indicator-based hunting, product-specific patching, or named-group attribution: it supplies no actionable campaign indicators, affected versions, or named actor.^[0]

Defensive actions

• Extend hardening coverage beyond a small set of designated high-value websites.: The reported campaign targeted 105 e-commerce sites, and the article explicitly connects successful low-cost attacks to insufficiently hardened sites. Its defensive analysis warns that protecting only selected crown-jewel sites leaves other exposed systems vulnerable.^[0]
• Address overprivileged accounts, reused credentials, and unmanaged service accounts, with particular attention to authentication and access controls.: Silverfort identifies these identity weaknesses as paths AI agents exploit. In its separate laboratory exercise, stronger authentication and access controls changed the outcome from rapid domain takeover to 21 dead ends and abandonment. This result supports the recommendation but is not evidence about the retail campaign's initial-access method.^[0]
• Automate access decisions where human handling would delay enforcement.: The report specifically recommends automated access decisions to prevent human response bottlenecks and emphasizes stopping attacks at the point of access. The eight-minute domain takeover was a laboratory result, but it illustrates the response-time pressure behind this recommendation.^[0]
• Prioritize protection and recovery of minimum viable business processes when allocating defensive investment.: Gambit Security argues that defending against every individual attempt becomes economically difficult when attackers operate at very low marginal cost. The report recommends identifying essential business processes, preventing their disruption, and measuring spending by its effect on business operations rather than by threat or tool counts.^[0]