Check Point Research March-April 2026 AI Threat Landscape Digest — single operator runs two AI platforms in parallel to breach nine Mexican government agencies; EvilTokens jailbreak-as-a-service

Unit 42 published a comprehensive write-up on Screening Serpens (a.k.a. UNC1549, Smoke Sandstorm, Nimbus Manticore) on 2026-05-22 covering operations from February through April 2026 timed to the onset of the U.S.–Israeli Middle East conflict that began 2026-02-28 (Unit 42, 2026-05-22 · Cybersecurity Dive, 2026-05-22). The group deployed new RAT variants across two malware families: MiniUpdate in four variants used between 2026-03-26 and 2026-04-17 with lures impersonating aviation, healthcare and financial-services firms, and MiniJunk V2 in two variants used between 2026-02-17 and 2026-03-27 against Middle Eastern and U.S. targets.

The technically significant evolution is AppDomainManager hijacking (T1574.014) paired with classic DLL sideloading (T1574.001): the infection chain drops a legitimate Microsoft .NET executable alongside a weaponised UpdateChecker.dll / InitInstall.dll / Updater.dll and — critically — a malicious .runtimeconfig.json that redirects the CLR's AppDomainManager loading at process startup, silently disabling ETW tracing and strong-name validation before the RAT executes. That leaves the host's EDR operating in a reduced-telemetry mode on every infected workstation. Delivery is high-touch — fake recruitment PDFs, spoofed video-conference meeting invitations, and ZIP archives containing a legitimate executable as the trigger; persistence uses scheduled tasks; C2 routes through Azure-hosted domains. Confirmed targets: U.S., Israel, UAE, plus at least two further Middle Eastern entities consistent with prior UNC1549 focus on aerospace, defence and telecommunications. The CH/EU nexus is indirect but real — Swiss aerospace and defence suppliers (RUAG, Pilatus and defence export channels) sit squarely in the sector profile, as do EU R&D firms historically swept up in Iranian collection campaigns.

Detection vantage: alert on .runtimeconfig.json writes by non-installer processes; watch the Microsoft-Windows-DotNETRuntime ETW provider for StrongNameVerification=0 startup events and CLR debug-mode initialisation; watch scheduled-task creation from processes with .dll parent images loading via rundll32.exe / svchost.exe. Hardening: enforce a code-integrity policy (UMCI + trusted-signers allowlist) so unsigned DLLs cannot load into the .NET CLR; restrict .runtimeconfig.json writes outside install paths via FIM.

Check Point Research's March-April 2026 AI Threat Landscape Digest (published 2026-05-22) is the operationally most striking annual / periodic AI report of the past month. The centrepiece — researched by Gambit Security and summarised in the Check Point post — documents a single unidentified operator compromising nine Mexican government agencies between December 2025 and February 2026, covering tax records, civil registry, patient files and electoral infrastructure. The structural innovation: the attacker ran two commercial AI platforms in parallel — one managing live exploitation and issuing >5,000 AI-executed commands, a second processing harvested data and feeding instructions back into the first. Persistence for the AI itself was simple: modifying the AI client's startup configuration file to embed persistent instructions inherited by every subsequent session.

Two further findings have direct EU/CH public-sector implications. First, the EvilTokens platform — a commercial jailbreak-as-a-service tool packaging AI-driven phishing generation, financial-data extraction and similar capabilities as a subscription — represents the same commoditisation curve as Kali365 (§ 1) but for AI-assisted intrusion. Second, CPR explicitly calls out that stolen API keys for Anthropic, OpenAI, Groq and Mistral are now high-value criminal targets, since they grant access to powerful AI services without an account; Swiss federal and cantonal agencies using commercial AI APIs should treat key rotation cadence and source-IP scoping (Conditional Access on the API layer) on par with classic privileged-credential hygiene. Detection vantage: bulk exfiltration events temporally co-located with anomalous API call patterns to commercial AI services from non-standard processes; process trees in which AI client libraries spawn data-collection subprocesses; cloud audit logs showing API key issuance followed immediately by large-volume inference calls from unusual source IPs.

HiddenLayer / Hadrian researchers disclosed CVE-2026-45829, a CVSS 4.0 = 10.0 pre-authentication RCE in ChromaDB's Python FastAPI server (affected from v1.0.0) (Hadrian Security, 2026-05-19; BleepingComputer, 2026-05-19). The vulnerable endpoint is POST /api/v2/tenants/{tenant}/databases/{db}/collections: when the request body sets trust_remote_code: true with an attacker-controlled HuggingFace model identifier (or a local path), the server fetches and executes the attacker-supplied Python code before the auth check fires, then politely returns 403 Forbidden after the code has run. The flaw exists only in the Python FastAPI server (chromadb[server] pip package) — the default Rust server (chroma run) does not traverse this code path. Per BleepingComputer's reporting of Shodan queries, approximately 73 % of internet-exposed ChromaDB instances are running a vulnerable version of the software. As of disclosure, ChromaDB v1.5.9 (latest) is unpatched. Mitigations: disable the Python FastAPI server and migrate to the Rust server; alternatively, block network-level access to the ChromaDB API (it should never be internet-exposed in the first place); if internal, set trust_remote_code: false server-wide via config. Detection concept — unexpected outbound network connections from ChromaDB Python server processes; child processes spawned by uvicorn / gunicorn workers with non-default lineage; access logs showing POST /api/v2/.../collections bodies referencing HuggingFace repository slugs with attacker-controlled patterns. T1190 Exploit Public-Facing Application; the impact maps to T1059.006 Python execution under the server context.

BigBlueButton (BBB) — the de facto open-source virtual classroom platform deployed across German DFN, Swiss SWITCH, and pan-European GÉANT academic networks, including cantonal school deployments — published three GitHub Security Advisories on 2026-05-17 covering distinct flaws in its bbb-web component, all in versions before 3.0.21 (two of three) and 3.0.23 (one). CVE-2026-46351 (CVSS 8.1) is a CWE-330 weakness: the sessionToken is generated with insufficiently random values, letting an authenticated low-privilege attacker who shares or has observed a meeting determine other participants' session tokens and impersonate any conference user (BBB GHSA-7959-pf2v-xc4h, 2026-05-17). CVE-2026-46353 (CVSS 8.1) is a CWE-284 access-control bypass in the presentationUploadExternalUrl endpoint: by supplying specific URL parameters an attacker can bypass checksum validation and send valid API requests to restricted endpoints without proper authentication, with high confidentiality + integrity impact (BBB GHSA-43hc-5g2m-cqff, 2026-05-17). CVE-2026-46404 (CVSS 6.8) is a CWE-918 SSRF in presentation URL validation: insufficient redirect-following checks allow a high-privilege authenticated attacker to reach RFC1918 and link-local (169.254.0.0/16) addresses from the BBB server context (BBB GHSA-xqm3-6q7q-4v5h, 2026-05-17). BSI's WID-SEC-2026-1568 corroborated on 2026-05-18 (BSI WID-SEC-2026-1568, 2026-05-18).

Why it matters to us: BBB is operated at scale by Swiss cantonal Volksschule deployments, German Länder ministries of education and university IT, EU national-research-and-education networks (NRENs). The combination of session-token prediction + checksum bypass would let a low-privilege classroom participant impersonate other students and teachers or send arbitrary authenticated API calls; SSRF on the server gives a presenter-role lateral-movement primitive into RFC1918 networks (KVM hosts, internal LDAP, SIS endpoints). Upgrade bbb-web to ≥ 3.0.21 for the first two CVEs and ≥ 3.0.23 for the SSRF; monitor bbb-web logs for anomalous joins using close-by sessionTokens and for API calls to presentationUploadExternalUrl carrying unexpected URL parameters; alert on egress from the BBB server process to RFC1918 / 169.254/16 ranges. MITRE T1212 (Exploitation for Credential Access) covers the session-token-prediction primitive; the SSRF maps to T1190 (Exploit Public-Facing Application) chained with internal-network reach.

UPDATE (originally covered 2026-05-13, 2026-05-15): Three concurrent developments show the TeamPCP / Shai-Hulud campaign has entered an open-source-imitator phase following Datadog Security Labs' 2026-05-13 analysis of the leaked Shai-Hulud worm source code. First, OX Security disclosed on 2026-05-17 four malicious npm packages published by deadcode09284814 — chalk-tempalte, @deadcode09284814/axios-util, axois-utils, and color-style-utils — combined weekly downloads ~3,000 (OX Security, 2026-05-17; The Hacker News, 2026-05-18). chalk-tempalte is a near-unmodified clone of the leaked Shai-Hulud worm with a modified C2 server and a new attacker-controlled key embedded in the code — the two primary sources disagree on whether this is a public or private key (see § 7); axois-utils bundles "Phantom Bot," a Golang HTTP/TCP/UDP/Reset-flood DDoS tool with Windows Startup folder and Linux scheduled-task persistence that survives package removal; the other two harvest SSH keys, cloud-provider credentials (AWS/GCP/Azure), and cryptocurrency wallet data.

Second, SANS ISC synthesised a 2026-05-18 campaign update confirming that Checkmarx officially acknowledged on 2026-05-11 that its Jenkins AST Scanner plugin had been trojanised — version 2026.5.09, compromise window 2026-05-09 01:25 UTC to 2026-05-10 08:47 UTC — making this TeamPCP's third confirmed Checkmarx intrusion in three months (SANS Internet Storm Center, 2026-05-18; Checkmarx, 2026-05-12). Hundreds of Jenkins controllers installed the malicious plugin before removal; remediated builds 2.0.13-848 and 2.0.13-847 are safe. CxSAST on-premise was unaffected; the cloud-integrated checkmarx/ast-github-action, checkmarx/kics-github-action, and VS Code extensions were all trojaned.

Third, SentinelLabs disclosed on 2026-05-07 — also folded into the SANS ISC summary — "PCPJack," a rival cloud worm that scans for exposed Docker, Kubernetes, Redis, MongoDB and RayML services and chains five CVEs (CVE-2025-29927 Next.js middleware auth bypass; CVE-2025-55182 Next.js Server Actions deserialization; CVE-2026-1357 WPVivid arbitrary file upload; CVE-2025-9501 W3 Total Cache RCE; CVE-2025-48703 CentOS Web Panel command injection) for initial access, then explicitly kills TeamPCP processes and removes TeamPCP artefacts before harvesting credentials — assessed by SentinelLabs with moderate confidence as possibly a former TeamPCP affiliate. Defender takeaway for the Swiss/EU public-sector SOC: developer endpoints and CI/CD runners with installed Checkmarx plugin should be audited for plugin versions outside the known-safe SHA range during the 2026-05-09 → 2026-05-10 window; npm audit and SBOM scans should flag the deadcode09284814 author/scope; egress from CI runners to *.lhr.life hostnames is a high-fidelity hunt pivot for the npm worm wave; Docker/Kubernetes/Redis/MongoDB endpoints exposed to the internet should be inventoried and removed from public exposure (PCPJack's scan list). MITRE T1195.002 (Supply Chain Compromise), T1552.001 (Credentials in Files), T1041 (Exfiltration over C2 Channel).

UPDATE (originally covered 2026-W21 weekly): VulnCheck honeypot telemetry confirmed active exploitation of CVE-2026-42945 on 2026-05-17, promoting the 18-year-old ngx_http_rewrite_module heap buffer overflow from PoC-public status (where it sat last week) to actively-exploited. The flaw is reachable by an unauthenticated remote attacker via a single crafted HTTP request to any NGINX instance running a rewrite-rule configuration that uses unnamed PCRE captures ($1, $2); successful exploitation crashes the worker process (DoS reliable on ASLR-enabled hosts) and reaches RCE on hosts where ASLR is disabled.

Affected per F5 PSIRT advisory K000161019: NGINX Open Source 0.6.27 through 1.30.0 (every release since 2008) and NGINX Plus R32 through R36, plus F5 NGINX Instance Manager, NGINX Ingress Controller, NGINX Gateway Fabric, NGINX App Protect WAF, F5 WAF for NGINX, and NGINX App Protect DoS. Patches: NGINX Open Source 1.30.1 / 1.31.0; NGINX Plus R32 P6, R36 P4. Interim mitigation if immediate upgrade is not possible: convert unnamed PCRE captures in all rewrite directives to named captures ((?P<name>...) syntax). Detection-engineering anchors that follow from the flaw class (heap-overflow worker crash under specific rewrite-rule configurations) are NGINX worker-process crash events (SIGSEGV / SIGABRT and immediate respawn) in syslog / journald, correlated with inbound HTTP requests carrying unusually long or deeply-nested rewrite-rule input strings from the same source; defenders should validate these against their own rewrite-rule configuration before depending on them.

As of 2026-05-14, Check Point published full analysis of the leaked 16.22 GB "Rocket" database. Administrator zeta88 announced a communications-infrastructure overhaul (new Tor addresses, new affiliate channels) rather than shutdown — the operator is actively hardening against exposure rather than exiting. Bedrock Safeguard's decryptor covered the pre-patch binary; the operator has claimed to patch the binary. Continued victim activity is expected. No new victim disclosures or Tor-address confirmations surfaced in W21 research; watch for new DLS address announcement.

Check Point's April 2026 monthly threat report (published early May 2026) confirms Qilin / Agenda leading all ransomware operators with 15% of 707 published attacks in April; Germany is the third-most-targeted country globally at 5.0% of victims (US 41.6%); Europe accounts for 27% of ransomware victims globally. Sector targeting in April 2026: Business Services (33.8%), healthcare, manufacturing. The Gentlemen — despite the May 4 backend breach — remained in the top-7 operators with 320+ victims (Check Point Research, 2026-05-08). The synthesis the dailies did not yet absorb: Germany's 5% share of global ransomware victims is materially elevated compared to the 2024–2025 baseline (~2–3%); the Qilin DLS lists 65 German victims total as of 2026-05-16 (Check Point blog, dataset reference). For Swiss defenders: CH-DE cross-border operations (Swiss subsidiaries in DE, German subsidiaries of Swiss parents) inherit the German exposure level; this is the empirical basis for a DACH-region threat-modelling premium on ransomware-readiness exercises.

UPDATE (originally covered 2026-05-10 in the Q1 2026 ransomware quarterly synthesis): Check Point Research published "Thus Spoke…The Gentlemen" on 2026-05-13, a detailed analysis of a 44.4 MB extract from the group's leaked "Rocket" backend database (16.22 GB total) that was posted to the cybercrime forum Breached on 2026-05-04 after the group's infrastructure was compromised by an unidentified actor (Check Point Research, 2026-05-13; BankInfoSecurity, 2026-05-11). The dataset contains 8,200 lines of internal chat-tool traffic across channels INFO / general / TOOLS / PODBOR, shadow files with password hashes, affiliate negotiation transcripts, and configuration artefacts for the ZeroPulse C2 framework.

Nine operator handles are identified — including administrator zeta88 (also hastalamuerte), who both manages the RaaS panel and participates directly in encryption events. Reconstructed attack chain: initial access almost exclusively via unpatched edge devices — FortiGate CVE-2024-55591 (the group's documented mainstay), Cisco appliances, CWMP/TR-069 interfaces — or purchased infostealer credentials; post-access tooling includes NetExec, RelayKing (NTLM relay), CertiHound (AD Certificate Services abuse), TaskHound, PrivHound; EDR-suppression utilities EDRStartupHinder, gfreeze and glinker manipulate ETW callbacks and NTDLL syscall tables; persistence is maintained via Cloudflare Zero Trust tunnels and self-provisioned WireGuard/OpenVPN chains.

Two operationally critical facts: (1) Check Point Research attributes a count of 1,570+ victim entries to a separately-exposed SystemBC C&C server, against 332 victims publicly listed on the group's data-leak site in the first five months of 2026 — significant under-reporting of true scope (Check Point's wider comparison cites 412 cumulative DLS listings); (2) the decryptor has been released as GitHub Bedrock-Safeguard/gentlemen-decryptor, enabling existing victims to recover without payment (decryptor disclosed in BankInfoSecurity's 2026-05-11 reporting). For Swiss / EU SOCs handling an active Gentlemen incident the workflow changes today: attempt decryption before any negotiation. Detection pivots from the leak: alert on EDRStartupHinder, gfreeze, glinker process names (custom binaries, not commodity); monitor for AD Certificate Services reconnaissance (certutil enumeration of CA servers and templates) consistent with CertiHound; correlate with FortiGate CVE-2024-55591 initial-access exploitation patterns that the group continues to weaponise.

UPDATE (TeamPCP / mini-shai-hulud first covered 2026-05-07; PCPJack worm covered 2026-05-10; this is a distinct new artefact in the same actor ecosystem): On 2026-05-09–10 (UTC) TeamPCP (UNC6780) published a backdoored build of the Checkmarx Jenkins AST plugin (version 2026.5.09, marketed under the actor's signature naming "Checkmarx-Fully-Hacked-by-TeamPCP") to the Jenkins Marketplace. Any Jenkins instance configured to auto-update the AST plugin during that window pulled the malicious build and executed the SANDCLOCK credential stealer in the runner context (Checkmarx — Ongoing Security Updates, last updated 2026-05-09; The Hacker News, 2026-05-11; SecurityWeek, 2026-05-11).

SANDCLOCK targets every secret reachable from a typical CI/CD pipeline environment: GitHub Personal Access Tokens, AWS / Azure / GCP credentials, Kubernetes service-account tokens, Docker / OCI registry credentials, SSH keys, and Checkmarx One API tokens. Affected pipelines should be treated as full secrets-compromise events: every credential the runner could read must be rotated and any artefact built or deployed in the window audited. Checkmarx's ongoing-security-updates page specifies plugin version 2.0.13-829.vc72453fa_1c16 (published December 2025) as the safe pinned version; a CVE has been issued as CVE-2026-33634 per the Checkmarx advisory. This is the third Checkmarx-product supply-chain compromise by this actor in three months, after the March 2026 KICS Docker image and the April 2026 VS Code extension defacement — the cadence and the actor's naming convention indicate persistent targeting of the Checkmarx product line specifically, not opportunistic distribution-channel abuse.

Mapped to T1195.002 Compromise Software Supply Chain and T1552.001 Credentials In Files. The GTIG AI Threat Tracker (see § 5) attributes SANDCLOCK specifically to TeamPCP and flags the stealer as explicitly designed to harvest LLM API keys in addition to traditional cloud credentials — consistent with the actor's pivot to monetising stolen LLM access. Defender pivot: inventory every Jenkins plugin auto-update enabled across CI/CD estates; constrain runners to short-lived OIDC-federated credentials (no long-lived PATs in runner env) where the platform supports it; audit Checkmarx One API logs for unexpected source IPs since 2026-05-09.

For every Jenkins controller running the Checkmarx Jenkins AST plugin: confirm installed plugin version; if 2026.5.09 was ever pulled (auto-update enabled, or manual install in window), declare a secrets-compromise incident, rotate every credential the runner could read (GitHub PATs, AWS / Azure / GCP access keys, Kubernetes service-account tokens, Docker registry credentials, SSH keys, Checkmarx One API tokens, and any LLM API keys exposed to CI), and audit any artefact built or deployed in the window. Pin the plugin to 2.0.13-829.vc72453fa_1c16 per Checkmarx's ongoing-security-updates page. Where the Jenkins platform supports it, migrate runners to OIDC-federated short-lived credentials so the next supply-chain compromise yields no usable secrets.