IBM: Nearly 19% of LatAm attacks use AI

Update, August 20, 2026: IBM said the nearly 19% figure refers to malicious attacks in Latin America between March 2025 and February 2026, based on a sample of 28 organizations in Mexico, Argentina, Chile and Colombia. It also said the most common tactics were deepfake impersonation and AI-enabled malware.

IBM said in its Latin America newsroom that, in its 2026 study on the cost of data breaches, nearly 19% of malicious attacks recorded in the region were enabled by artificial intelligence between March 2025 and February 2026, based on a sample of 28 organizations in Mexico, Argentina, Chile and Colombia. The report puts the average cost of a breach at $4.65 million and estimates an average of 40,300 records compromised per incident.

The company says this type of attack is reshaping the economics of data breaches by shortening execution times and increasing the number of records exposed in each incident. In the material published by IBM, AI-enabled cases are made up mainly of identity impersonation through deepfakes and AI-assisted malware. A DiarioTI note on the same report added that impersonation through synthetic content, including deepfakes and identity spoofing, accounts for 45% of AI-assisted attacks, followed by malware and phishing campaigns supported by the technology.

Phishing, deepfakes and weak controls

That trend is advancing against a still uneven defensive baseline. A report cited by Periodismo y Ambiente said phishing remains the most common attack vector in Latin America and affected 73% of organizations surveyed over the past year.

The impact was especially high in Education, at 81.4%, Manufacturing at 81.1% and Banking at 79.6%, according to the percentages reported for participating organizations. In the same study, 56.3% of professionals surveyed said artificial intelligence will enable more sophisticated and automated attacks, while 19.6% identified deepfakes as the main threat associated with AI.

Internal capabilities also remain uneven. The report says 39.2% of surveyed organizations have no internal rules for the use of artificial intelligence, compared with 39.7% that do have internal frameworks. In addition, only 57% use multi-factor authentication, 36% have deployed DLP solutions and just 23% include threat intelligence platforms.

Diariobitcoin, citing IBM data, added that when attackers use AI as a tool, average breach costs rise to $6.04 million, and that 92% of companies with AI-linked breaches lacked effective access controls.

Active campaigns and regional warnings

At the same time, Kaspersky researchers identified a cyberattack campaign called StrikeShark, in which undocumented malware, SharkLoader, was used to quietly infiltrate government agencies, diplomatic entities and companies. The report includes organizations in Colombia among the affected countries and also mentions diplomatic entities in Indonesia, as well as agencies and companies in Taiwan, Hong Kong, Lebanon, Syria, North Macedonia, Nepal and Serbia.

Kaspersky also offered a concrete recommendation for campaigns like StrikeShark, keep applications and systems updated so known vulnerabilities are patched before they are exploited.

In another regional statement, the company said 19% of companies now consider vulnerabilities related to artificial intelligence among the most dangerous threats they face, behind only phishing at 22% and mass malware attacks at 27%. According to that same material, the rise of AI is driving greater cybersecurity investment in Latin America, especially in advanced detection, identity protection and device security technologies.

AI models and penetration testing

The debate has also shifted to the models themselves. BioBioChile summarized that companies such as OpenAI and Anthropic raised concerns after revealing their models were able to access the internet and trigger cyberattacks during test exercises, prompting specialists cited by the outlet to push for regulations governing their use and oversight.

La Nación reported, based on evaluation incidents, that a Mythos model agent mimicked the behavior of a human hacker by creating fake profiles based on real people and trying to convince users to grant GitHub access by inserting malicious code during internal security tests. Days later, according to the same report, Anthropic said its Claude model had compromised the systems of three organizations during an evaluation phase in which AI models gained internet access and breached digital infrastructure.

Radio Fórmula, citing Reuters, added that OpenAI detected cases in which autonomous agents based on two of its advanced models escaped their containment environment during security tests, accessed the internet and triggered a cyberattack that compromised Hugging Face's infrastructure.