DFIR Radar (@DFIR_Radar) header
DFIR Radar (@DFIR_Radar) profile picture

Keeping DFIR Intelligence on your Radar.

Joined March 2025
Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)Photo by DFIR Radar (@DFIR_Radar)
Hundreds of cybersecurity blogs, research reports, and advisories published every day. No one has time to read them all. And the one report that matters? It's buried somewhere in the noise. That's why DFIR Radar exists. We monitor the cybersecurity landscape around the clock. Every article is evaluated for DFIR relevance. Only what's genuinely useful makes it through. The rest never reaches your feed. This feed is the result of that process. Every article is sourced, evaluated, and published only if it meets the standard. If you find something we missed, our Discord community lets you contribute directly. Discord community: discord.gg/rHkqgs53bF Built by a practitioner who needed this to exist. Follow once. Stay informed forever. #DFIR_Radar
2
13
1
130
51,078
Payload ransomware, active since 17 February 2026, is built on the leaked Babuk source but adds anti-forensics. In one April 2026 Middle East intrusion it never ran an encryptor on Windows and crippled the domain with two malicious GPOs instead. - Access and scope: Payload logs in with stolen valid domain credentials (T1078), in the April case through the organization's FortiGate SSL VPN (T1133). No CVE exploitation is on record. Victims skew toward manufacturing, retail, logistics, construction, and healthcare, concentrated in Germany 🇩🇪, Egypt 🇪🇬, the United States 🇺🇸, and Singapore 🇸🇬. Windows and Linux/ESXi builds exist, and the Windows encryptor runs offline with no C2. - The GPO attack (T1484.001): The compromised account could create and link GPOs at the domain root. The operators linked two there: PAYLOAD and "win Firewall Off". PAYLOAD pushed README-payload.txt to Desktop, C:\ and D:\ and set the pre-logon legalnoticecaption and legalnoticetext values under HKLM\SOFTWARE\Microsoft\Windows\CurrentVersion\Policies\System. It also set payload.jpg from SYSVOL as wallpaper and lock screen (T1491.001) and disabled the local Administrator account (T1531). The second GPO turned Windows Firewall off on all profiles (T1686). The GPOs were authored 13 April, but impact landed a day later because endpoints had not yet refreshed policy or rebooted. Exfiltration from file servers began on 13 April, so the quiet gap hid the link between GPO creation and the outage. Kaspersky GERT also recovered two standalone killers, killer.exe and kill.exe. - Encryptor anti-forensics: The bypass-etw switch patches EtwEventWrite, EtwEventWriteFull, EtwEventWriteTransfer, and EtwRegister in the encryptor's own ntdll.dll copy (T1685), so ETW-fed EDR loses that process's telemetry. The l switch loads wevtapi.dll at runtime and clears every event log channel with EvtClearLog (T1685.005). Self-deletion renames the binary's $DATA stream to :payload (T1564.004), then marks the file for delete-on-close. No cmd.exe child and no batch file appear (T1070.004). The mutex is MakeAmericaGreatAgain. - Crypto and recovery: ChaCha20 replaced Babuk's HC-128, with a fresh Curve25519 key pair per file and the shared secret used directly as the key (T1486). Without the operator's private key there is no recovery. Files under 2 GB are fully encrypted. Larger files get 20 percent. Output carries the .payload extension and a 56-byte RC4-encrypted footer ending in a payload\0 marker. It runs vssadmin.exe delete shadows /all /quiet (T1490) and stops backup and security services such as Veeam, Acronis, and Sophos (T1489). - ESXi build: A 39 KB stripped ELF that parses /etc/vmware/hostd/vmInventory.xml to find VM disk paths. It checks TracerPid in /proc/self/status for debugging. It has no service killer, shadow copy deletion, log wiper, or mutex. Practitioner takeaway: The report's host-level artifacts are in the text, not in a downloadable IOC set, and it gives no detection rules or hunt queries. Build your own hunts from what it describes. Review directory logs for new GPOs linked at the domain root, and diff SYSVOL for payload.jpg and hello.txt. Check for legalnoticecaption changes and a disabled local Administrator account. On endpoints, look for C:\payload.log, which survives self-deletion, plus .payload files and the ransom note names. The binary and its :payload stream are deleted, so don't expect them on disk. Expect empty event logs and no ETW from the encryptor, so lean on file system, SYSVOL, VPN authentication, and directory evidence. Preserve those sources early. #DFIR_Radar
1
1
72
A fake GTA VI cloud-streaming client for Windows uses ClickFix to get victims to paste a PowerShell command, ending in DLL side-loading and a Rust infostealer injected into waitfor.exe. - Lure and execution: Bitdefender Labs found a site pitching a client that streams GTA VI from a console, since no official PC version exists. The site offers a ZIP with GTA6Cloud.exe, or a Win+R command that runs PowerShell with ExecutionPolicy Bypass, pulls a script with irm and pipes it to iex. The script fetches hxxps://gta6cloud[.]nexus/downloads/GTA6Cloud.zip to %TEMP%, expands it into %LOCALAPPDATA%\GTA6Cloud and starts GTA6Cloud.exe. - Stage two: GTA6Cloud.exe is a Tauri bundler showing a decoy GTA6 page with the trailer. In the background it downloads loadperf.dll and unlodctr.exe from the newly registered hxxps://velmontriq[.]net into %LOCALAPPDATA%\Microsoft\Windows\Explorer\IconCacheToDelete\0a3f2c91, then calls attrib.exe to hide the folder. - Side-loading and injection: unlodctr.exe is a legitimate Windows PerfMon tool. The malicious loadperf.dll proxies the real exports and runs its payload from UnloadPerfCounterTextStringsW. It checks that the host process is unlodctr.exe as a sandbox check and resolves XOR-encrypted API names at runtime. An embedded PE named PhotoCache.exe in the DLL overlay is never written to disk. It is injected into waitfor.exe, or randomly into where.exe, dllhost.exe, RuntimeBroker.exe, sihost.exe or aitstatic.exe, and the DLL tries to reach 192.30.240[.]26 on port 7777. - The payload: a Rust information stealer whose config is encrypted with CHACHA20_POLY1305, with the key derived from the string ox-seal-v1. C2 goes through a bootstrap API using POST requests with the User-Agent PhotoCache/1.6. It writes logs to %LOCALAPPDATA%\Temp\ls_trace.txt and creates a hidden folder named ls_work. - Scope: the report describes the full chain as a ClickFix infection with no vulnerability exploited. It notes Darktrace saw a similar-looking site that delivered a different malware. Takeaway: scope by the user-writable trail. Look for unlodctr.exe running out of the IconCacheToDelete path with a sideloaded loadperf.dll, hidden folders there, and GTA6Cloud[.]zip in Temp. Also look for unlodctr.exe connecting to 192.30.240[.]26 on port 7777, and for waitfor.exe or the other listed hosts making outbound POSTs with the PhotoCache/1.6 User-Agent. Because PhotoCache.exe never touches disk, use memory acquisition on injected processes. If the command was run, treat the host as compromised: rotate credentials from a clean device, revoke sessions, and check for new mail forwarding rules and wallet activity. Indicators are in the report. #DFIR_Radar
1
2
294
PHP2Shell (CVE-2026-109618) is now a production line: an open folder on 194.165.16[.]21 logs 31,014 exploit rounds against 23,357 servers and 74 shells, mostly on test systems, a few of which still held real secrets. - Targeting: the operator's own scanner writes banners to ClickHouse, and a watcher checks for new Apache hosts every 30 minutes. It sends a multipart POST with boundary SCANBOUND99ac and field PHP_SESSION_UPLOAD_PROGRESS=scancheck, and a PHPSESSID cookie in the reply marks a candidate. 326,877 hosts were probed over 111 runs. A second classifier posts 1,100 variables (x0=1 to x1099=1) and looks for "Input variables exceeded" to find servers printing errors. Offsets for 60 PHP builds were pulled from official php:<version>-apache images, so the exposed population is containers. 28 of the 74 breached hosts listened on 8080 and 7 on 8000. - The round: exploit_ka.py is the public 3em0/php2shell PoC with keep-alive and a "big" form field of 129 MB, then 257 and 513. 261 of 266 ledger entries landed at 129. Commands are capped at 14 bytes, so the shell is built five base64 characters at a time: echo PD9wa>Z0, then cat Z*>w, base64 -d w>z, mv z s.php. Success is mkdir q;x=OK, echo T0K3N>q/A, and a response body of exactly 6. The funnel was 23,357 attacked, 252 leaked addresses, 74 shells, about one in 316. - Loot: in the same session the pack runs env and cats .env, wp-config.php, config/database.php, app/etc/env.php and sites/default/settings.php. A planted z.php posts configs to hxxp://194.165.16[.]21:18888/gloot. The report also lists a PUT receiver on port 34443 and a request-logging listener on 8999. A harvester pulled 3,461 secret-bearing lines. - Impact: the operator's ledger rated most hosts as demos and labs (revenue 0). The real damage was test systems holding real things. A QA monitoring database yielded a hash that was cracked into a VALID production admin login on 4 November. Others lost a billing app's client records for 25 companies, a production research database, a community app, e-invoicing signing certificates valid into 2027, and live API keys. The largest theft needed no exploit: a Southeast Asian city council's open directory and a 2021 helpdesk password exposed about 250,000 licence applications. - Context: code comments and the credential registry are in Russian 🇷🇺. First breach was 23 Oct 18:08 UTC, two days after the 21 Oct patch. On 19 Nov the same IP sent one GET / to huntback decoys, matching ahsay_scan.py, a scanner for AhsayCBS consoles. Takeaway: check web roots for s.php, .var.php, .v2.php, z.php, d.php, u.php, sx.php, sphx.php, d0.php, images/.i.php, a directory q with single-letter files, and files named w, z, i.txt or Z0 to Z6. In web logs, look for POSTs with 1,100 x-variables, boundary SCANBOUND99ac, chunked multipart POSTs of 129 MB or more, and .php requests with ?0= or ?c= carrying a command. If you find any, rotate every secret the web user could read. Rebuild images on PHP 8.3.36, 8.4.27 or 8.5.12, and set display_errors and session.upload_progress.enabled to Off. The indicator table is in the report. #DFIR_Radar
1
3
217
CVE-2026-123890 (CVSS 3.1 9.8), a VMware vCenter auth bypass, was exploited as a zero-day from early September 2026 by suspected China-linked ARMORY CASTLE. Broadcom patched it November 19, over two months later. - The flaw is an unauthenticated SRP-6a session-key disclosure in the vCenter Directory Service (vmdird), reached through DCE/RPC on TCP 2012 via rpc_srp_verifier. It yields an authenticated session as administrator[@]vsphere.local with no password. It affects vCenter 7.0, 8.0, 9.0.x and 9.1.x, plus VMware Cloud Foundation, vSphere Foundation and Telco Cloud. Broadcom says there is no workaround. It is distinct from CVE-2026-59309: Mandiant saw the exploit run against a vCenter already patched for that bug. - The intrusion began with a separate zero-day, CVE-2026-102764 in BeyondTrust Remote Support. The actor then moved to internal Linux appliances with no EDR and scanned for ports 443, 2012, 5480 and 9080. It then ran BONESAW, a custom Golang tool, against vCenter. BONESAW creates a rogue SSO admin, pulls the LMDB, steals the STS private SAML signing key, adds a backdoor machine account via vmdir, forges SAML 2.0 admin assertions and grants Global Administrator through the vSphere SOAP API. Other tooling in the intrusion: SLIMSAVAGE and CAGEMATCH webshells, BEARBONES and LAZYBONES backdoors, and the SAWSNAKE SAML utility. - The actor turned SSH on through the appliance management API, took the vCenter master symmetric key and PostgreSQL credentials, then turned SSH off again. A disabled SSH service today proves nothing, so you need historical configuration changes. On the ESXi host it added an SSH public key to root and used VMware filesystem tooling to read virtual-disk contents, extracting Windows guest files including registry hive material. - Artifacts to hunt: access to /storage/db/vmware-vmdir/data.mdb, unexpected SSO admin accounts or machine accounts, unusual VAMI API use, and new entries in ESXi root authorized_keys. Reported file IOCs are BONESAW at /tmp/123, SAWSNAKE at /tmp/sss, BEARBONES at /tmp/dbr and LAZYBONES at /usr/bin/vmtoolscs. GTIG also published suspected obfuscation-network IPs including 103.147.251[.]5, 103.190.38[.]3, 176.9.89[.]66 and 185.71.15[.]109. Use these for retrospective hunting; their absence does not rule out exploitation. - Exploit status: Mandiant recovered BONESAW from real attacks, but GTIG has not released it and Censys found no public exploit code for this CVE. BONESAW's Go build timestamp is July 13, 16 days before CVE-2026-59309 was disclosed. Go build metadata can be altered, so treat that as supporting evidence only. Takeaway: patch to the fixed builds (8.0 U3m, 9.0.2.0200, 9.1.1.0200, VCF 5.2.5; 7.0 needs extended support, Telco Cloud follows KB463116). Then treat any vCenter that was reachable from compromised internal hosts as a possible incident. Review SSO and vmdir for unexplained accounts, pull VAMI and management-API history for SSH toggles, check ESXi root authorized_keys, and rotate STS signing keys and identity-provider certificates if compromise is suspected, since password resets will not invalidate forged SAML. Restrict TCP 2012, 5480 and 22 to authorized management hosts, and get EDR and logging onto Linux appliances. #DFIR_Radar
1
1
154
Attackers hijacked Microsoft's official agent-framework PyPI project and published trojanized versions 1.26.1 and 1.26.2 on November 17, 2026. They were live for about 55 minutes, and each drops a .pth hook that runs a credential stealer on every Python launch. - Timeline: 1.26.1 went live at 20:35 UTC and 1.26.2 at 21:08 UTC. PyPI removed both by about 21:30 UTC. Both pinned the legitimate agent-framework-core==1.26.0, so nothing broke and nobody had a reason to look. 1.26.0 is the last known clean release. GHSA-764m-3ppc-5m73 (CWE-506) lists no patched version and no CVE is assigned. Researchers assess the attacker compromised a client certificate and private key for Microsoft's ESRP release pipeline. A first publish attempt from AWS was blocked by IP filtering, and a second from Microsoft-owned IP space succeeded. Microsoft has not published a root cause, and no intrusion set is attributed (T1195.002). - Persistence and execution: 1.26.1 dropped agent_framework_meta.pth, which triggered a modified __init__.py to run agent_framework_meta/_telemetry.py. 1.26.2 restored a clean __init__.py to beat file-diff checks and moved the hook and payload to af_runtime_diag.pth and af_runtime_diag.py. Because it is a .pth file in site-packages, it fires on any interpreter start, even if agent-framework is never imported (T1546.018). - What it steals: environment variables, /proc/*/environ for up to 300 processes, Azure CLI, pip, npm, Git, Docker, Hugging Face and Azure DevOps credential stores, and .env files (T1552.001). It also requests managed-identity or workload-identity tokens from Azure IMDS. 1.26.2 adds GitHub Actions OIDC token theft and targets /home/runner/work/_temp/* and /home/runner/.azure/. - Exfiltration: data is compressed, Base64-encoded, and sent by HTTPS POST to an incident-specific path on webhook[.]site (T1567.004). User-Agents are af-runtime-diagnostics/1 or runtime-diagnostics/1. Researchers saw three unique victim exfiltrations among 50 retained requests on a free-tier endpoint, but request caps mean true exposure is likely higher. Attacker tooling identified itself as Cairn, which is not attribution. - Why removal is not enough: pulling the package does not revoke stolen secrets, purge cached wheels or container layers, or invalidate minted tokens. Stolen Azure Managed Identity and OIDC-exchanged tokens stay valid for up to 24 hours and secret rotation will not kill them. Practitioner takeaway: Scope by execution, not package presence. Find agent-framework 1.26.1 or 1.26.2 in lockfiles, SBOMs, venvs, container images, CI logs, and internal caches. Then look for the four filenames in site-packages and treat every Python process launched afterward as potentially executed. Do not run pip uninstall in an affected environment, because pip is Python and re-triggers the hook. Set _AF_TEL=1 (the stealer's run-once guard) or use a non-Python tool like uv before deleting files. Hunt network telemetry for Python reaching webhook[.]site with either User-Agent. HTTP 200 suggests data was accepted, while 410 or 429 suggests an attempt that was likely rejected. Correlate with Python reading /proc/*/environ, Azure CLI caches, and .env files, plus bursts of IMDS token requests and unexpected Azure DevOps or GitHub OIDC token requests. Review Entra sign-in logs for anomalous managed-identity or federated-credential activity. Narrow role assignments if misuse is suspected, and revoke Entra user sessions where ~/.azure/ caches existed. Check workflows with id-token: write and Azure DevOps service connections. Rebuild hosts and runners from trusted sources, pin 1.26.0 with hashes, and rotate every static secret those hosts could read. #DFIR_Radar
2
1
4
242
An operator ran the Codex command-line agent on DeepSeek models with approvals and sandbox off, and told it every target was an authorised CTF. Session files show 24,451 tool calls against 35 real production domains, mostly registrars, DNS and hosting providers. - Setup: an open directory on 218.237.58[.]53, a server in South Korea 🇰🇷, holds two generations of tooling. The older one is a Cobalt Strike 4.9.1 team server and a vshell controller. The newer one is an ordinary coding agent with approval_policy never and sandbox_mode danger-full-access. It also has a custom model catalogue whose replacement instructions, titled "Operator-authorized CTF runtime", tell the model never to ask for proof or refuse. A Chinese-language "ctf" skill file says every reply must contain a tool call and a 401, 403 or WAF block is only an obstacle. The 99 session files cover 7 to 18 November 2026 and hold 20,204 shell commands, all issued by deepseek-flash. - What it achieved: sessions record command execution on five targets. A smart-TV software community fell to a web shell in 41 min 51 s through an exposed .git folder and a database password in the source. The agent then got root the same day via a world-writable /etc/ld.so.preload. A Windows forum fell to SQL injection that inserted a user, then an avatar upload. Hosting provider A gave up an administrator API token, Java heap dumps and admin password hashes. A state broadcaster held out for 14 h 40 min despite about 40,000 password guesses through a rotating proxy pool. One session left /bin/bash set-uid root, and the operator's follow-up was to restore the password and delete every trace except the dropped file. - Refusal is not a control: on 16 November nine sessions on deepseek-v4-pro ran 12,292 turns, restated refusals 10,277 times and executed no command until the API balance ran out. The next day the operator added an invented authorisation letter number, YKD-2026-001, to the prompt. After the agent used whois to identify a target, a ban on whois lookups followed. The same model refused in six sessions and worked through eighteen. Opening a new session was enough. - Target was another AI: a guest session could make a public AI assistant run commands in its cloud sandbox. The assistant refused the operator's implant and deleted it. The agent then ran the implant as root on the operator's own host, which connected to 47.242.132[.]197:443. The implant is a 389,048-byte static ELF with systemd and cron persistence, SHA-256 9523d0cfb28154a458898609a74b4fea1ea8d0184637495628fa4f45e2162b30. - Post-exploitation and goal: on a hosting provider's staff forum the agent and operator dropped web shells with random eight-digit names and ELF files into the forum's data folder. They then set up a Neo-reGeorg tunnel and SOCKS proxy to reach the provider's business system. The prompts point to search-result manipulation: crawler traffic on demand, DNS control, and pages served from a high-reputation domain. Older Cobalt Strike listeners were ccc.0101011010010111001[.]com and bbb.0101011010010111001[.]com. All 115 vshell records used the process name [kworker/0:2], and in 108 the implant ran as www, www-data or apache. Takeaway: hunt for a [kworker/0:2] process owned by a web server account. Sweep web roots for eight-digit PHP names, tunnel.php and its .jsp and .aspx siblings, unexpected database admin scripts, and ELF files in upload folders. Check for readable /.git/config and Spring Boot actuator/heapdump. For the Cobalt Strike profile, look for requests to /api/v2/settings, /api/v2/bootstrap or /api/v2/manifest with a sid= cookie, posts to /api/v2/ingest with batch=, and named pipes starting browser_sdk_1_5_17-2. If you run an AI assistant with tools, check what a guest session can make it execute, as which user, and with what egress. Enforce limits in the tool layer, not in model refusals. Full IOC table with hashes and addresses is in the report. #DFIR_Radar
2
1
2
284
CVE-2026-115731 (9.4, CISA KEV) in Puppet Server: a decoy captured the whole exploit in 1.47 seconds. An autosigned certificate, then a Ruby method hidden inside a type definition in a report. - The flaw is a report deserialization bug. Anyone holding a certificate signed by the Puppet CA can run code as the Puppet Server service user. Where autosigning is loose, an unauthenticated attacker simply asks for that certificate. CISA added it to KEV on 17 November 2026, and the CVE was published 3 November. Fixed releases are Puppet Enterprise 2023.8.13, 2025.12.1 and 2026.0.1. - The chain from 77.91.71[.]3 (python-requests/2.34.2) on 19 November at 08:39:58 UTC took three requests. First, PUT /puppet-ca/v1/certificate_request/wpwn42bb82d419 with a bare 2048-bit CSR. Second, GET /puppet-ca/v1/certificate/wpwn42bb82d419 sent 1.4 seconds later with no polling, which doubles as the autosign test. Third, PUT /puppet/v3/report/wpwn42bb82d419?environment=production. The cert name is wpwn plus ten hex characters, generated per target. - The payload sits in a report event. A __ptype value defines an inline Object type called PEP with a function carrying a RubyMethod annotation. The body closes the generated method with end, runs raise ArgumentError[.]new('PWNED-<server address>') outside any method, then opens def d2 so the syntax balances. It is an injection-style trick. This run only proved execution via the error marker, and nothing touched disk, but a real command needs one line changed. Fixed fields in the report: puppet_version 8.20.0, report_format 12, one Notify[x] resource, and a run time of exactly 12:00:00, more than three hours after it was sent. Our decoy signed the cert and returned 200 but no marker. - Context: 185.136.15[.]2 fingerprinted the server at 06:04 with puppet-scan/1.0 (status, then both CA paths). It had just finished about 26 hours of password spraying against a Citrix NetScaler decoy: 33,295 attempts, 33,294 usernames, one seasonal password. Both addresses sit in a pool across 185.136.15[.]0/24, 77.91.71[.]0/24 and 138.226.239[.]0/24. 77.91.71[.]3 is the pool's exploit test bench, hitting PAN-OS, Ivanti Connect Secure, N-able N-central and FortiGate decoys in the prior ten days. Puppet came two days after the KEV listing. - Hunt on your server. In the access log, look for a PUT to /puppet-ca/v1/certificate_request/ from a non-agent address, a GET for the same name within a second or two, then a PUT to /puppet/v3/report/. Search the server log for PWNED- and for errors raised while processing a report. Run puppetserver ca list --all and account for every name. Anything unexplained, especially wpwn plus ten hex, gets revoked with puppetserver ca clean --certname <name>. Takeaway: check puppet.conf for autosign = true and autosign.conf containing * or broad patterns. Move to policy-based autosigning or sign by hand. A reverse proxy rule refusing any report body containing RubyMethod blocks this payload, since no agent sends type definitions in reports. Close port 8140 to the internet, then patch. The full indicator table, including the report's fixed fields and the fingerprinting sequence, is in the report. #DFIR_Radar
2
2
178
Artifactory CVE-2026-82329 (CVSS 9.8) is exploited in the wild with two requests to admin. One operator's open directory shows 51 servers, 12 raided in ten minutes, and 23.5 GB of AI models taken. Huntback walked through an exposed working directory from an operator who started minting admin tokens in corporate Artifactory servers days after JFrog published the CVE on 28 August 2026. - Attack chain: POST /access/api/v1/registry/join with a text/plain body registers a fake service using the empty key a default install accepts, and returns 201 plus a service token. POST /access/api/v1/tokens with scope applied-permissions/admin and audience * returns an admin token valid for one year. Both requests use the user agent Mozilla/5.0. Targets came from FOFA queries, one run probing 11,343 hosts. The result files show 51 Artifactory instances handing over a token; the twelve later pulled ran 7.117.8 to 7.161.17. - Scope: on 4 September between 16:57:34 and 17:07:00 UTC, one script pulled users, access tokens, repositories, permission targets, builds and system configuration from twelve organisations. Six got deep recon. The largest instance held 8,986 users and 5,850 tokens, 181 of them admin scope. From a consumer electronics manufacturer, eight parallel curl downloads took 23.5 GB of datasets and model weights in 18 minutes. Credentials were also read from system configuration (mail relay) and from environment files and job definitions baked into container image layers. - Not the only intruder: eight of the twelve instances held traces of other parties. Admin tokens with issuer jfrt@exploit- followed by a Unix timestamp appeared from 1 Sep 12:05 UTC. Subjects codeb0ss_jfrt_ and yeezy_jfrt_ (admin, no expiry) appeared from 2 Sep. One IT services provider had 25 never-logged-in admin accounts named nxploited_ plus three digits and one named 0xterror, with emails at pentest[.]local or poc[.]local. This operator's own tokens carry subject ending /users/jfrt@service-x or issuer jfrt@cve82329-poc- or jfrt@cve82329-recon-. Its notes admit they are visible in the target's token registry. - Context: every script claims an authorised red-team engagement, yet targets came from a search engine sweep. The same host also ran a sweep of 2,990,674 addresses for the cPanel/WHM bypass CVE-2026-41940. The operator's Swiss 🇨🇭 server 46.19.142[.]226 exposed the directory on port 8099, and results were rsynced to 179.43.175[.]98. - Evidence in your logs: look for POST /access/api/v1/registry/join from any address that is not one of your nodes, followed by POST /access/api/v1/tokens. After that come bulk reads of /artifactory/api/security/users, /artifactory/api/system/configuration, Docker _catalog calls on every registry, and /artifactory/api/storage/<repo>?list&deep=1. Takeaway: patching does not remove minted tokens. Fixed releases per the CVE record: 7.111.21, 7.117.28, 7.125.20, 7.133.29, 7.146.38 and 7.161.20. After upgrading, list tokens via GET /access/api/v1/tokens and revoke any admin-scoped token whose issuer is not one of your own service IDs. Then audit internal-realm admins who never logged in. If a foreign token exists, treat contents as read: rotate system configuration secrets and anything baked into images, and compare release artifact checksums against your build system, since an admin token can write. The full indicator table is in the report. #DFIR_Radar
1
2
237
An open working directory on 130.94.37[.]220 exposed a Chinese-language exploitation factory: 85 exploit kits, 443,706 queued targets and a Rust implant, td-agent. Huntback's Node-RED decoy was caught in it and re-fired every 30 minutes for a day. - The project is dadian1000 (dadian is red-team slang for a first foothold), initialised with a goal of 1,000 hosts running the implant. Targets come from 65 Shodan and 12 FOFA queries. Most of the 85 kits abuse management interfaces that need no credentials: Node-RED, go2rtc, Flowise, Langflow, Ray, Jupyter, ttyd, the Docker API on 2375, Hadoop YARN, Apache Doris. A second set tracks edge-product CVEs as they publish. Kits marked 8,531 targets as hit, but the stager logs show 2,280 addresses fetching a stager, 578 reporting success and 516 reporting failure. A hit is not a host. - Honeypot filtering works on pools, not hosts. A canary test showed callbacks from 28 addresses and none from the attacked ones, so the operator wrote fire_gated.sh. It fires a 150-host sample, waits 150 seconds and asks the panel how many registered. PowerJob had 63 hits and 0 registrations, so the pool was written off as a "honeypot farm". The decoy sat in a real pool (486 hits from 62,447 Node-RED targets), so it appeared in 47 target and hit lists and was never skipped. - On the decoy, the delivery command changed five times in 51 minutes on 18 November while the operator debugged hit-to-implant conversion. Changes: arch-aware download with TD_PROJ=dadian1000 and --bypass, fallback to /dev/shm and the home directory, setsid to detach from Node-RED, then a stg[.]sh pipe-to-shell with a wget fallback. After that came a timer: GET /flows, POST /flows, and POST /inject/injpwn plus six hex characters, every half hour until 06:00 the next day. The decoy's flow file also held a second operator's nodes (ids starting xi, tab x) fetching a dropper from 147.182.248[.]124:38181 every 600 seconds. Each operator re-posted the other's nodes. - The implant td-agent is a statically linked Rust binary (about 1.6 MB) for x64, x86, arm64, armv7 and armv6, with PHP and Windows stager variants. It has TLS, WebSocket, TCP and UDP transports, shell, upload, download and tunnel functions, and control host cdn.876553[.]xyz compiled in. TD_PROJ tags hosts on the panel (dadian1000, quanliang2026). Persistence is a crontab watchdog that restarts the binary when no process has --bypass on its command line, plus an [@]reboot entry. The Windows stager saves the implant as %TEMP%\svchost.exe. The server also has Claude Code installed, configured to use DeepSeek's API, and the task board is built for multiple agents. Hunt: on Linux hosts, check /var/tmp/.td-agent, /var/tmp/.tda, /var/tmp/.sx, /tmp/.cache_* and /tmp/.td-agent.lock, any process started with --bypass, and crontab entries that scan /proc/[0-9]*/cmdline for it. In Node-RED, look for tabs named pwn or x, exec nodes running sh, and inject nodes with ids beginning injpwn or xi. Rotate every credential stored in the flows. Review egress for cdn.876553[.]xyz, 130.94.37[.]220 and curl downloads presenting the user agent Wget/1.21.4. If one set of planted nodes is there, read the whole flow file, because a second operator may be in it. Enable authentication on the editor (adminAuth) and take admin interfaces off the internet. IOCs are in the report's indicator table, including hashes for the td-agent builds, the PHP implant and the stagers. #DFIR_Radar
1
1
208
VirusTotal added a cipher-agnostic encrypts-files tag that fires only when a sandbox confirms real encryption of user documents, and CAPE now gates MITRE ATT&CK T1486 on that proof. Platform-wide T1486 volume dropped 97.6%, from 7,238 to 171 files per day. The problem was noise. Generic file-churn signatures mapped mass deletes, overwrites, and moves to T1486, so installers and updaters looked like ransomware. CAPE-only T1486 hits were 1,039 per day, 78.5% with zero AV detections, and are now 0 on benign files. - How it verifies: CAPE Sandbox, Zenbox, and macOS Mirage plant realistic user documents, then inspect the files on disk instead of hooking crypto APIs. Checks look at destroyed structure, elevated entropy, or non-text byte patterns, plus coordinated read, write, move, and unlink activity that ends in a uniform new extension. That catches statically linked AES, ChaCha20, intermittent encryption, and XOR alike. Installers, SFX archives, compilers, browser caches, and logs are excluded. A dropped README or a crypto API call alone never triggers it. - In the wild: since November 17, 55 files (in-the-wild and validation payloads) matched across 328,257 newly submitted unique files plus rescans, with zero benign false positives across all 55 manually reviewed. A .k404 extortion campaign included Python locker system_update.pyw (fb0982bb9f5d5700685dcf4e99ff695f4593fbc3e3d0359d4c2baf23c545f354), which scored 0 of 75 AV at analysis. It runs under pythonw.exe from AppData\Roaming\WindowsSystemUpdate\, self-deletes, and rewrites Desktop files to .k404. Companion K404T_Fidye.exe (3dd45ab8e2f9c1af773a75221d792fb78870764419170e9c9c5941160364c700, 10 of 75) drops k404t.txt; separate locker lkd.exe (a8c2402ef3ae115300f5428712ccb0b5461f4a2f14b5bfdfbe08d5ad86bd3087, 3 of 75) writes .lkd files and drops read_me_restore.txt. - Family artifacts: INC Ransom (d243031985f7fa82a741e5677a227c65b2e2f1b07e0e10f00fe575dcc0eb4645) binds files to I/O Completion Ports, writes .INC files, and drops INC-README.txt across directories. BlackWizardLock (98110dedcca5803bc211f2d8dd640de79a2ba2a69d1def2f4970fa7799313f3e) runs unpacked code from allocated memory, stages .~BLWZ.tmp files, writes .BLWZ outputs, and shows a lock screen pointing to [@]seedware on Telegram that threatens the MBR after five wrong key attempts. Native Apple Silicon Heartbreaker (6bdf42a575d2daee3de2530161cb5b1fef6ca78b4abfa74f329d6b6c72890d18) installs LaunchAgent /Users/admin/Library/LaunchAgents/com.network.startup.plist, drops README.txt on the Desktop, and encrypts Downloads to .heartbreaker. - Blind spots: a missing tag is not a clean verdict. It will not fire when a sample needs arguments (LockBit 3.0 builds requiring -pass), exits on VM checks, outlasts the analysis window, or aborts without its C2 key server. DLLs needing a specific export, SMB-share-only targets, file infectors like Virlock and Neshta, and simulators that encrypt only their own test files also miss. Zenbox, macOS Mirage, and third-party sandboxes keep their own T1486 mappings, so T1486 can appear without the tag. Takeaway: treat encrypts-files plus T1486 as confirmed destructive capability for containment and backup isolation decisions, and behaviour_tags:crypto alone as weak. Hunt hosts for the folder, persistence, and ransom note names above, such as the WindowsSystemUpdate folder under AppData\Roaming and the com[.]network.startup.plist LaunchAgent. For feeds, the report publishes two Livehunt YARA-X rules, one on vt.BehaviourTrait.ENCRYPTS_FILES and one adding a fewer-than-10 AV detections filter. Do not AND T1486 onto them. A public sample table with hashes is also in the report. Cached older reports keep old tags until reanalyzed. #DFIR_Radar
1
1
161