AI Threat Landscape Digest March-April 2026
During the March–April 2026 reporting period, AI use in offensive operations advanced from development and planning to real-time operational deployment. Multiple independent cases, involving individual criminal actors, mass exploitation platforms, ransomware groups, and state-sponsored espionage, show evidence of commercial AI models executing autonomous attack workflows across extended campaigns. Key findings: AI-orchestrated attacks have progressed from experimental, state-sponsored use to in-the-wild criminal deployment. Multiple criminal operations relied on commercial Claude Code as a persistent operational tool in multi-week campaigns. Agentic configuration files are being weaponized as persistent jailbreak vectors. Hooks, project-level files, and settings files abuse the operational control level and redefine the model behaviour at the architecture level. AI-enabled attack platforms are commercializing AI capabilities. Operators can now buy access to platforms where the AI pipeline, model selection, jailbreak, and delivery mechanisms are embedded in the product. AI provider credentials have become a high-value target. As commercial AI services become central to offensive operations, API keys for Anthropic, OpenAI, Groq, Mistral, and HuggingFace are harvested at scale from compromised .env files, providing access without registration and resilience against provider attempts to revoke this access. AI as Live Attack Operator AI selection considerations Underground forum discussions still show actors debating the use of commercial models, dedicated jailbreak services, or locally hosted open-source models, reflecting the lower-skill end of AI adoption. More advanced actors combine tools pragmatically: from commercial AI models, open or uncensored models where commercial providers restrict output, and custom automation pipelines that perform repetitive analysis at scale. Tasks are systematically broken down into smaller sub-requests that present a lower apparent risk profile. Forum users further discuss and share methods and alternatives to avoid mainstream-provider safety controls by mixing open-weight Chinese frontier models, privacy-routed proxies, and explicitly uncensored services. The Mexico Breach When Anthropic disclosed GTG-1002, a Chinese nexus campaign using Claude Code for cyber espionage, in November 2025, this was seen as an experimental, state-sponsored development. The disclosure carried no IoCs and was therefore disputed by independent researchers, and the activity was detected only through Anthropic’s own API monitoring. The Mexico breach, which occurred a few months later, demonstrates similar architecture in operational, financially motivated criminal use, at scale, and with a recovered forensic record. Between late December 2025 and mid-February 2026, a single operator compromised nine Mexican government agencies. Researchers documented the case after recovering materials from attacker-controlled VPS servers. Details include the operational record: 1,088 attacker prompts generating 5,317 AI-executed commands across 34 sessions. The breach scope was significant: tax records, civil registry data, vehicle records, patient files, and electoral infrastructure were affected. However, an even more important lesson is how the campaign was run. The operator built a dual AI workflow. Claude Code served as the interactive exploitation assistant, helping advance access, write exploits, build tunnel chains, map victim environments, and escalate privileges. In parallel, harvested server data was processed through GPT-4.1 for automated intelligence analysis. The GPT output was then used to task new Claude sessions. As we highlighted in our previous review, the agentic infrastructure itself was exploited to bypass the model’s safety restrictions. At the start of the campaign, Claude refused to execute requests which it correctly identified as offensive cyber activity. The attacker then changed tactics. Instead of asking Claude to generate malicious content directly, they pasted a large penetration-testing cheatsheet into CLAUDE.md in the project root, the file Claude Code automatically loads as persistent project context at the start of every session. From that point on, subsequent sessions inherited the rules and techniques in that file. The attacker did not need to repeat the jailbreak as the behavior persisted through the project configuration layer. After gaining root on a civil registry server, the model’s actions in subsequent sessions were consistent with the persistent cheatsheet, including unprompted post-exploitation steps such as shadow file extraction and timestamp cleanup. Bissa Scanner A second documented case, Bissa Scanner, was published in April 2026, after researchers identified an exposed operator server. Bissa is a modular mass-exploitation platform built around React2Shell (CVE-2025-55182), with 900+ confirmed compromises across millions of scanned Next.js endpoints and an archive of 30,000+ distinct .env filenames recovered from operator-controlled S3 storage. The operation has been running since September 2025. Here, AI is positioned one step back from the exploitation layer: Claude Code and OpenClaw (running claude-sonnet-4-6, with a Telegram bot for triage alerting) served as the operator’s working environment for reading the scanner codebase, troubleshooting, refining the collection pipeline, and prioritizing high-value access. No jailbreak was documented and commercial Claude was accessed through the standard API. Bissa harvested .env files specifically for AI provider credentials (Anthropic, OpenAI, Groq, Mistral, OpenRouter, HuggingFace, Replicate, DeepSeek). AI provider credentials have become a deliberate target, valuable enough for sophisticated operators to enumerate and harvest at scale alongside conventional credential theft. These credentials are likely intended to be used in future offensive criminal activity and attribute it to the legitimate account holder instead of the attacker. Agentic Configuration Files: A Persistent Attack Surface The previous section demonstrates the use of agentic configuration files to override safety features in their own AI sessions. The same inheritance mechanism can be used in reverse: an attacker plants malicious agentic configuration files in a repository, and an innocent developer uses the project and becomes the next victim. A recent CPR report documented three exploitation paths and disclosed two (now patched) CVEs. CVE-2025-59536 exploits Claude Code’s Hooks feature (hooks, .claude/settings.json), executing arbitrary commands before the developer can read them. A parallel path uses .mcp.json to trigger the MCP server startup, bypassing the consent dialog entirely. CVE-2026-21852 redirects ANTHROPIC_BASE_URL to a malicious proxy that intercepts authorization headers and potentially steals API keys, granting read/write access to the entire team Workspace before any trust prompt appears. The attack vector in all three cases is “supply chain”, a malicious settings file embedded in a pull request, honeypot repository, or compromised codebase that results in system compromise on the developer machine. The underlying issue of using agentic configuration files as the attack surface and supply chain is not specific to Claude. The potential attack surface is architectural and may apply equally to Cursor (.cursorrules), Windsurf (.windsurfrules), and GitHub Copilot Workspace (.github/copilot-instructions.md). AI-Powered Fraud at Scale: EvilTokens EvilTokens represents a category of offensive tooling offered for sale: a commercial Phishing-as-a-Service (PhaaS) platform, built using AI and operating an LLM pipeline as a runtime component of the attack. A buyer with no AI knowledge can purchase access to a fully integrated pipeline in which model selection, jailbreak, and output delivery are handled at the platform level. EvilTokens runs a multi-stage attack flow. Device-code phishing pages impersonating Adobe, DocuSign, and SharePoint harvest Microsoft OAuth tokens. The AI pipeline then activates these tools: Via Groq, llama-3.1-8b-instant ingests up to 5,000 emails in 250-email batches, extracting account numbers, routing numbers, wire amounts, payment deadlines, and reporting hierarchies. Also via Groq, llama-3.3-70b-versatile synthesizes the intelligence, generates BEC (Business Email Compromise) drafts tailored to the victim’s writing style, and assigns a BEC score. gpt-4o-mini translates stolen emails for non-English-speaking operators. The SMTP Sender delivers the output with rotating SMTP pools, header fingerprint randomization, DKIM signing, and CSS randomization. The researchers assessed with high confidence that the platform’s backend was AI-generated. The model choices reflect deliberate task routing: Llama 3.1 8B was used for cheap high-volume extraction, Llama 3.3 70B for reasoning-heavy synthesis and stylistic mimicry, and GPT-4o-mini was reserved for translation where it has the strongest multilingual capability and where the task itself looks innocuous to provider-side monitoring. The riskiest content generation is kept on Groq-hosted open-weight models instead of on OpenAI’s more closely monitored surface. The jailbreak is the product. Both Groq-hosted LLaMA stages operate under a jailbreak embedded at the platform level, not applied by the operator and not visible to the customer. Stage 1 frames the model as an “authorized red team security analyst” conducting “sanctioned penetration tests”; Stage 2 upgrades to “senior red team analyst.” Prompts direct the model to reference real email threads, mask payment changes behind “plausible business reasons”, imitate sender style, and generate emails “realistic enough to fool a trained employee.” This is security bypass at SaaS scale: write the jailbreak once, ship it as a feature, and it’s inherited in every customer session. The original EvilTokens advertising posts reveal additional features, including a Calendar Invite module which sends fake meeting invitations that appear as legitimate Outlook and Gmail meeting requests, with built-in Sender Spoofing (Organizer Identity). In a BEC context, this is used to apply timing pressure on finance personnel: a fake “urgent review meeting” appears on the target’s calendar shortly before a wire-transfer request lends the request a sense of pre-authorized context. Combined with the AI-generated email and the SMTP Sender, this completes a full BEC social engineering toolkit covered end-to-end by a single PhaaS offering. EvilTokens’ Telegram channel announced additional AI-based features after Sekoia’s disclosure. The platform did not go offline and accelerated its AI feature development through April 2026. The Vulnerability Race: AI on Both Sides of the Patch Window AI-assisted vulnerability research has become a category in its own right and is now commercialized at both major frontier labs simultaneously on two tiers: a restricted research-grade capability and a productized defender tool. At the frontier, Anthropic’s Claude Mythos, released through Project Glasswing, reportedly demonstrated a systematic, rapid mechanism to search for vulnerabilities and revealed a very large number of vulnerabilities, some long-buried zero-days in core infrastructure. These include a 27-year-old OpenBSD TCP/SACK bug found at roughly $20,000 in compute, a 16-year-old FFmpeg H.264 codec flaw, and a FreeBSD NFS remote code execution vulnerability in software that was analyzed for decades. The capability jump within a single generation is steep: on the same Firefox test set, Opus 4.6 produced 2 successful exploits and Mythos produced 181. Anthropic notes that this capability was not explicitly trained for but “emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.” The productized tier is wider and more accessible: Claude Security (running on the public Opus 4.7 model) entered public beta for Enterprise customers, and OpenAI’s Codex Security, in research preview since early March, has had 14 CVEs assigned during the preview window on OpenSSH, GnuTLS, libssh, PHP, and Chromium. The same capability curve is reaching attackers at the commodity tier, faster than defenders can patch. A researcher using a standard Claude API subscription identified CVE-2026-34197, a 13-year-old Apache ActiveMQ remote code execution vulnerability, and attributed roughly 80% of the work to Claude and the remainder to his refinement. LMDeploy SSRF (CVE-2026-33626) was exploited within 12 hours of the advisory publication, with no public proof-of-concept available. This time-frame compression is consistent with attackers building working exploits directly from advisory text. GenAI is accelerating this workflow. Vendors are using AI to find vulnerabilities that sat undiscovered in core infrastructure for decades while attackers are using AI to find and weaponize newly-disclosed vulnerabilities within hours of publication. The patch window, the period between disclosure and exploitation, is being compressed on both sides. Vendors and customers need to adjust to a new high rate of patch development, delivery and deployment. The side that reacts the fastest will gain the most from recent AI developments. Enterprise Adoption and Exposure Corporate environment data collected by Check Point in March – April 2026 shows enterprise GenAI usage continuing to scale while the associated risk profile remains stable. Approximately one in every 28 prompts (3.6%) posed a high risk of sensitive data exposure, a modest increase from the January–February baseline of 3.2%, observed across 91% of organizations actively using GenAI tools (compared with 90% in the previous period). The proportion of prompts containing potentially sensitive information rose from 16% to 18%. The average employee generated 78 prompts during March – April, up from 69, with organizations using an average of 10 GenAI tools. Interaction volume is rising while risk ratios remain stable, producing a proportional increase in absolute exposure events. The consistency of these metrics across two reporting periods indicates a maturing adoption pattern: data exposure is not an episodic incident category but a continuous operational risk requiring sustained monitoring and policy enforcement. Conclusion Our findings converge on a small number of structural observations. AI now operates as an attack component, not just as a development aid. The Mexican breach illustrates this at government-breach scale, and Bissa at mass-exploitation scale. The same commercial Claude Code architecture appears independently across criminal operations with different motivations and geographies, and in state-sponsored espionage. The convergence is operational consensus, not coincidence. The techniques aren’t new but the performance envelope is. Network scanning, credential spraying, lateral movement, BEC drafting, and vulnerability research all predate AI. What’s changed is the speed (working exploits generated from advisory text alone within 12 hours of disclosure), scale (one operator reaching the operational footprint of an advanced team), and breadth of knowledge (cross-domain expertise on demand lowers the entry requirement for sophisticated multi-vector campaigns). Defences calibrated to human attack tempo and human team throughput are not equipped for the AI equivalents. The AI attribution gap is structural. All the operations we documented in this report were discovered through attacker OPSEC failures or LLM provider monitoring, not through victim-side controls. AI-executed commands resemble skilled human activity closely enough to evade current behavioral controls. Operations that do not fail at OPSEC, or that route through stolen credentials or self-hosted models, remain unclassified. “The Turkish Rat” Evolved Adwind in a Massive Ongoing Phishing Campaign Check Point Research Publications August 11, 2017 “The Next WannaCry” Vulnerability is Here Check Point Research Publications March 12, 2026 “Handala Hack” – Unveiling Group’s Modus Operandi SUBSCRIBE TO CYBER INTELLIGENCE REPORTS We value your privacy! BFSI uses cookies on this site. We use cookies to enable faster and easier experience for you. By continuing to visit this website you agree to our use of cookies.
research.checkpoint.comMay 26, 2026extracted