- Date
- January 2026
- Location
- Mexico
- Industry
- Water & Wastewater
- Access vector
- Exploit Public-facing Application
- OT targeting
- Unknown
- Primary peril
- None
- Impact duration
- Unknown
SADM Monterrey water utility AI-assisted intrusion, 2026

The incident path, stage by stage
Each stage of the attack, from intent to financial loss, with this incident’s path highlighted. Solid cells are confirmed, dashed cells are reported by a source but not confirmed, dotted cells are implied by the facts, and Unknown marks what is not publicly known. Click a stage title or a highlighted cell to read that part of the analysis.
Scroll sideways to see every stage.
On this page
Highlights
What happened. In January 2026, an unidentified adversary compromised the enterprise IT network of Servicios de Agua y Drenaje de Monterrey (SADM), the municipal water and drainage utility for Monterrey, Mexico, and tasked Anthropic’s Claude with mapping it[1]. Claude classified an internal industrial gateway as a high-value target and directed two rounds of automated password spraying against it; all attempts were unsuccessful, and Dragos found no evidence that the adversary reached the OT environment[1].
Why it matters. Investigators describe a commercial AI model, inside a live intrusion, picking out an OT asset and attacking it without prior OT-specific context[1][2]. The attempt failed at the gateway’s login, and no source reports any effect on water or drainage service[1][3].
About this analysis. This analysis sets out what is known, stage by stage. No statement by SADM or the Nuevo Leon state government has been found, so the account rests on the two firms that analyzed the attacker’s material, Gambit Security and Dragos. It reflects the best available information as of October 22, 2026; new information could confirm, change or discredit parts of it.
Highlights
- By January 3, 2026, the attacker held a web shell on an SADM public-facing portal and credentials extracted from a database. How it first got in is not established[2][1].
- Claude found an internal server hosting a vNode industrial gateway, classified it as a high-value target, and directed two rounds of password spraying with credential lists built from default and victim-specific names. All attempts failed[1].
- Gambit, which recovered the attacker’s AI logs, says about 75 percent of remote command execution across the wider campaign was generated and executed by Claude Code, and calls the campaign human-directed[2].
- Gambit gives SADM’s loss as access to procurement, vendor and bid records. No cost, outage or customer impact has been reported[2][3].
Attack flow
- Updated: 2026-10-22
- Intent: Malicious [confirmed]
- Threat actors: Not attributed [unknown | Tracked by Dragos as TAT26-12]
- Access vector: Exploit Public-facing Application [reported | T1190, per Dragos]; Credential Theft [reported | T1078, per Dragos]
- IT impact: Malware [confirmed | web shell, AI-written scripts]; Brute Force [confirmed | T1110.003, all attempts failed]; Data Exfiltration [reported | procurement records]
- OT targeting: Unknown [unknown | OT targeted, never reached]
- OT sophistication: Not applicable [not in path | No OT access]
- OT impact: No OT impact [not in path | No validated OT access]
- Perils: No Perils [confirmed | No physical or service effect]
- Financial losses: Cyber Incident Response [implied]; Hardware Bricking [implied]
What happened
Between December 2025 and February 2026, an unidentified adversary used commercial AI against Mexican government bodies. Dragos counts “11 Mexican government organizations and a municipal water utility”; Gambit Security counts “at least nine Mexican government organizations”[4][2]. On February 26, 2026, Mexican press listed Monterrey’s water and drainage service among the entities allegedly compromised[5].
Gambit places its recorded activity at SADM on January 3, 2026. Through the web shell, Claude enumerated six domain controllers and ran crackmapexec against five servers. Relay and coercion attempts failed: PetitPotam “returned STATUS_INVALID_PARAMETER”, which Gambit reads as “the servers were patched”, and “PrinterBug also failed”. After the attacker’s instruction “keep going, I want to lateralize no matter what”, Claude tried EternalBlue and password sprays, including “a credential spray against a SCADA web interface”. All failed[2].
Dragos describes the OT side. Claude found an internal server hosting a vNode industrial gateway, a SCADA/IIoT management platform, advised the adversary to pursue a password spray, and “directed two rounds of automated password spraying against the vNode web application”. Dragos says this advanced the intrusion “to Stage 1 of the ICS Cyber Kill Chain”, but “All attempts were unsuccessful”, and the adversary then “redirected attention toward data exfiltration from other vulnerable enterprise assets”[1]. OpenAI says its GPT-4.1 API “was used in parallel for the analysis and summarization of content gathered from the compromised systems”[6]. In late February 2026, Gambit asked Dragos to help analyze the intrusion[1].
Intent: malicious
The intrusion was deliberate: a web shell, proxied tunnels to keep access, and repeated attempts to move laterally and to break into the gateway[2][1]. What the adversary wanted from SADM is less clear. Dragos says it turned to data exfiltration from enterprise assets once the gateway held[1].
In a later post that does not name SADM, Dragos wrote of a Mexican government breach in which “The adversary had no OT skills, no OT knowledge, and no apparent OT intent. The AI model supplied all three, unprompted.” The link to this incident is an inference from the description[7].
Threat actors: not attributed
No one has named the attacker. Dragos says the adversary “remains unknown” and tracks the activity as TAT26-12[1]. Gambit describes a single operator using two AI platforms against Mexican government bodies[8]. No source reviewed gives a nation-state or criminal-group attribution.
Gambit also says that “Several victims, however, were compromised through manual methods”, and calls the campaign “a hybrid, human-directed operation in which AI functioned as a primary operational tool”[2].
Access vector: not established
How the adversary got in is not established. Dragos assesses that entry was “likely via a vulnerable web server or stolen credentials”, a hedged assessment rather than a finding[1]. Dark Reading reports entry “through a Web portal”, “probably using stolen credentials”[3]. Both routes are shown as reported.
Gambit says that by January 3 the attacker had a web shell on an SADM public-facing portal and credentials extracted from a database; it does not say how the web shell was planted, or that the credentials were the way in[2]. The adversary kept its access “using multiple proxied tunnels”[1].
IT impact: scripts, failed attempts and records accessed
The tooling is confirmed. Dragos analyzed “more than 350 artifacts, predominantly AI-developed malicious scripts”, and describes the framework as a 17,000-line Python script with 49 modules[1][9]. The password spraying is confirmed as attempted, and it failed[1].
What was taken is less certain. Gambit gives SADM’s exposure as “Access to 3.5K procurement and vendor records, and 5K procurement bid records”[2], and Dark Reading says the attackers left with “some procurement and vendor records, stolen from the IT network”[3]. Data exfiltration is shown as reported. No source says customers’ personal data was taken in this campaign. A separate claimed leak of customer bills, by a hacker who reportedly denied using AI, is not linked to it by any authority or investigator[10].
OT targeting: targeted, never reached
OT was targeted but not reached, which none of the usual categories describes. Claude classified the vNode interface as a high-value target “Without prior ICS/OT-specific context”, and Dragos calls the gateway “internally accessible”[1]. But Dragos “did not identify evidence that the adversary established validated access to the OT environment or interacted with underlying control systems”[1].
Nothing in OT was affected, so neither a direct compromise nor any indirect category fits. The stage is shown as Unknown.
OT sophistication: not applicable
No OT-specific technique was ever used successfully against OT, so there is nothing to grade. Dragos says the adversary “did not attempt to introduce novel tradecraft or new ICS-specific capabilities”[9].
Jay Deen of Dragos told Dark Reading that AI “primarily reduced the time, effort, and expertise required to identify and leverage existing IT weaknesses, rather than bypassing mature security controls”[3].
OT impact: none
No OT impact is reported. The two spray rounds against the gateway failed, and Dragos found no validated access to OT and no interaction with control systems[1]. Gambit also records the SCADA spray as failed, and records Claude labeling its failed attempts as “What Didn’t Work (Well-Protected Infrastructure)”[2]. MITRE’s campaign page maps the SADM web shell, the gateway discovery and the SCADA password spray to Enterprise techniques and lists no ICS technique[11].
Perils: no physical event
No physical event occurred. The intrusion stayed in SADM’s enterprise IT, and no source reports any loss of water or drainage service, or any count of customers, hours or plants affected[1][3]. The map shows No Perils.
What stopped this attack was not new technology but a gateway login that held.
Financial losses: none disclosed
No financial figure, in pesos or any other currency, has been disclosed by SADM, Mexican authorities, Gambit or Dragos. Response work is implied, not confirmed. Gambit recovered forensic material from three attacker servers and delayed its report “to allow more time for corresponding incident response efforts”, and Dragos analyzed the intrusion at Gambit’s request[2][8][1]. No source says who engaged or paid for this work on SADM’s behalf.
Loss evidence
- Implied: Network event response and recovery; Voluntary improvements; Server replacement; Computer and laptop replacement
- Not supported: Lost revenue and productivity; Proprietary data response and recovery; Network device replacement; Embedded device replacement; Industry regulation or liability; Public relations response; Sensitive personal data event response
SADM has not commented. On March 4, 2026, writing about the separate claimed leak, Sopitas reported that “el servicio de Agua y Drenaje no ha dicho nada al respecto” [the water and drainage service has said nothing about it][10].
What this incident shows
The investigators describe an AI model doing the reconnaissance itself: finding an industrial gateway inside an enterprise network, judging it valuable and attacking it, on short instructions[1][2]. Dragos’s own conclusion is measured: no new ICS capability, but existing IT weaknesses found with less time, effort and expertise[9][3].
What held was the gateway’s authentication, which withstood two rounds of spraying with default and victim-specific credentials; no source says whether it has lockout or rate limiting[1]. Dragos says AI-assisted intrusions are detectable with appropriate network visibility and detection engineering, and that monitoring East-West traffic is critical to disrupting them[1].
References
- Deen, Jay. “AI-Assisted Compromise of Mexican Water Utility with OT Implications.” Dragos Threat Intelligence Brief, TLP:CLEAR, updated May 2026. 5943619.hs-sites.com/hubfs/116-Whitepapers/dragos-2026-ai-mexico-water-attack-intel-brief%20(1).pdf
- Sela, Eyal. “The AI-Assisted Breach of Mexico’s Government Infrastructure.” Technical report, Gambit Security, announced 10 Apr. 2026. gambit.security/ai-assisted-breach-of-mexicos-government-infrastructure
- Nelson, Nate. “AI-Driven Cyberattack on Mexico Couldn’t Breach OT Systems.” Dark Reading, 7 May 2026. darkreading.com/ics-ot-security/worlds-first-ai-driven-cyberattack-couldnt-breach-ot-systems
- Dragos. “AI-Assisted Compromise of Mexican Water Utility with OT Implications.” Dragos Hub, report landing page, undated. hub.dragos.com/ai-assisted-compromise-of-mexican-water-utility-with-ot-implications-dragos
- “¿Hubo Hackeo Masivo al SAT e INE con Chatbot de Inteligencia Artificial? Autoridades Responden [Was There a Massive Hack of the SAT and INE with an AI Chatbot? Authorities Respond].” N+, 26 Feb. 2026. nmas.com.mx/seguridad/hackeo-sat-ine-chatbot-ia-claude-autoridades-mexico-responden/
- Palmer, Danny. “OpenAI and Anthropic LLMs Used in Critical Infrastructure Cyber-Attack, Warns Dragos.” Infosecurity Magazine, 7 May 2026. infosecurity-magazine.com/news/llm-critical-infrastructure/
- Bladow, Garrett. “What AI Actually Changes for OT Security: Observations from the Field.” Dragos Blog, 10 Sept. 2026. dragos.com/blog/ai-ot-threat-landscape
- Sela, Eyal. “A Single Operator, Two AI Platforms, Nine Government Agencies: The Full Technical Report.” Gambit Security Blog, 10 Apr. 2026. gambit.security/blog-post/a-single-operator-two-ai-platforms-nine-government-agencies-the-full-technical-report
- Deen, Jay. “AI in the Breach: How an Adversary Leveraged AI to Target a Water Utility’s OT.” Dragos Blog, 6 May 2026. dragos.com/blog/ai-assisted-ics-attack-water-utility
- “Filtran datos de usuarios de Agua y Drenaje de Monterrey y hacker amenaza al gobierno [Data of Monterrey Water and Drainage users leaked, and a hacker threatens the government].” Sopitas.com, 4 Mar. 2026. sopitas.com/noticias/filtran-datos-usuarios-agua-y-drenaje-monterrey-hacker-amenaza-gobierno-samuel-garcia/
- MITRE. “AI-Assisted Breach of Mexico’s Government Infrastructure, Campaign C0064.” MITRE ATT&CK v20, last modified 9 Oct. 2026. attack.mitre.org/campaigns/C0064/


