Pinned Tweet
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
Trigram impersonates about 50 AI and blockchain brands to trick developers into running an infostealer themselves. Stormshield assesses with high confidence it continues the UNK_DeadDrop cluster, tied to North Korea-linked activity.
- Lure and gating: Recruitment, conference, and package-tracking lures, a supply-chain compromise, and trojanized repos all converge on the same infrastructure and payload. The fake recruitment site fingerprints visitors, geofences by IP (possibly to exclude China 🇨🇳), and demands a referral code, which suggests pre-identified targets. Rejected codes return "Already used," "Locked to another IP," or "Invalid code." Accepted visitors are told to paste a "technical check" command into a Linux terminal. That first command only POSTs a nonce, downloads nothing, and the page replies "Operation timeout. Please try again. Nonce Refreshed." Then the Copy button is changed to inject the real command into the clipboard, even if the user presses Ctrl+C.
- Loaders and payload: Windows runs PowerShell to fetch v.cmd, Linux runs bash to fetch a file named u, and macOS runs bash to fetch verify_m. These prep the environment: a Base64-encoded script sets the working directory, creates a lock to block concurrent installs, and checks for Node.js 18 or later, installing Node.js 20.18.1 if absent. They then fetch install_guard.js, which deploys an infostealer that collects system metadata, browser credentials, crypto wallets, and sensitive files. C2 is resolved through a GitLab dead drop.
- Repo hijack: On October 31, 2026, a forged commit to the antd-img-crop GitHub repo added a prebuild hook, prebuild.js, that fires when a developer runs npm run build. The npm package itself was not impacted. The chain installs Node.js via a fake installer, runs an encrypted next stage in memory, and ends with a third stage that steals browser passwords, cookies, keychains, wallets, SSH keys, and .env files on macOS, Linux, and Windows, then installs the Overlord RAT disguised as Google software. On macOS and Linux a fake dialog asks for the system password to "apply system updates and verify drivers and security settings." Same payload also arrives via a fake Amber Group careers portal and a sideloaded VS Code or Cursor extension named google-dev-tools.google-update-support, likely launched through a coding assessment and deleted after running.
- Back office: Stormshield found a management layer on 170.205.30[.]27:8000 running a platform called rms-platform. API routes show it scrapes GitHub profiles, sends phishing, manages decoys, issues access codes, exports victim data, and monitors crypto wallets. Visitors are rated important, moderate, or useless. It leans on GitHub and GitLab for hosting, Brevo, Resend, Mailgun, and Smartlead for email, and FullStory for session recording. Decoys and consoles ship in the same JavaScript bundle, with hostname or URL path choosing the interface.
- Attribution: Links to UNK_DeadDrop and Kimsuky include the console super-admin eros1030109@gmail[.]com, which matches an address used in Kimsuky's Operation GitPower. Also noted are asteara[.]org, the MongoDB username silviattech1129_db_user, and the Telegram recruiter account pichi_finance_hr. Operation GitPower itself used local LLMs via Ollama and GPT4All to build decoys, LNK files that run PowerShell to pull lures from GitHub raw content, and AsyncRAT. It has since targeted Ukrainian 🇺🇦 diplomacy and policy experts, mostly unsuccessfully per Genians. SafeDep cautions that its Contagious Interview match rests on technique and does not prove who the operators are. Jamf found a Zoom-themed Overlord variant sharing the com.zoom LaunchAgent label and com.zoom.plist name with FlexibleFerret, but no direct overlap with UNK_DeadDrop.
Takeaway: Start on developer workstations. Look for terminal or PowerShell history showing a pasted download-and-run of v.cmd, u, or verify_m, unexpected Node.js 20.18.1 installs, install_guard.js on disk, and node spawned from a prebuild.js hook during npm run build. Check VS Code and Cursor for google-dev-tools.google-update-support (it may have been deleted, so look for traces), and hosts for a remote-access agent disguised as Google software. If a host ran any of it, treat SSH keys, .env secrets, browser sessions, and wallets as exposed. Review repos for forged commits touching build hooks. The attribution indicators and infrastructure details are in Stormshield's report.
#DFIR_Radar
1
45
Source: thehackernews.com/2026/11/tr…
9
An exposed workspace on 38.76.169[.]207 leaked a CyberStrikeAI conversation database. It ties one 16 November 2026 operator session to model accounting, process events, stored tool executions and agent-recorded findings. Approval enforcement was switched off mid-session, and the agent's final report claims more than the evidence supports.
- The evidence chain reconciles. The report found 221 stored tool executions against 240 process-level tool calls. The 19 local results with no execution ID (ten skill actions, four task-plan updates, three file reads, two exit actions) explain the gap. All 221 execution IDs match the stored table. The conversation says agent_mode supervisor, but all 240 tool-call events name the agent as main and the role as orchestrator, so there is no evidence of a worker swarm.
- The approval gate failed, then was disabled. Four approval interrupts were all rejected by audit_agent with the comment "LLM call failed; conservatively reject". These were audit-agent model-call failures (the review call has a 90 second timeout), not human decisions. Three ended at 90 seconds. The fourth ended at 43 seconds when the user cancelled the run. At 09:48:15 UTC the saved conversation configuration shows enabled=0, mode=off, and the recovered source returns without requiring approval when config is absent or disabled. The user later approved proposed actions in chat at 10:18, which is wording, not an enforced per-call gate. The database does not say who changed the setting, and the report found no evidence the model exploited the approval system.
- What the agent actually did: it consumed a handoff document, rechecked old access paths, ran network reconnaissance, then focused on a compute-marketplace MCP interface. The c2_listener tool returned zero on both calls and the webshell list showed no connections. Of the 221 executions, 170 were the execute tool (160 completed, 10 failed). The strongest supported result is returned compute-node and job data. No shell, internal-network access, data-lake breach or destructive change is established.
- Where the final report overreaches: the nine findings are agent-created, not nine validated vulnerabilities. An anonymous token that returned a JWT and key was a throwaway identity, not a stolen one. A job query total of 10000 with relation "gte" is a lower bound, not 10,000 downloaded records. The "260-node dump" came from a script that looped 26 offsets and appended ten hits each time with no unique-ID check, while nearby results report a node-search total of 259. A test a minute earlier returned the same first node at offset 10. Distinct-node count cannot be computed from the recovered files.
- Accounting: eight usage rows name deepseek-v4-pro, with 211 model calls and 13,564,316 total tokens (13,353,943 prompt, 210,373 completion). These are local application records, not an invoice, and the prompt total may include repeated context.
Takeaway: investigate audit-agent failures, configuration changes and subsequent tool execution as one sequence, and preserve the previous config and the actor where the app logs them. Do not accept claims of execution, access or impact without process, file, identity or transaction evidence, and keep response bodies or protected hashes, because request logs cannot show volume retrieved. Also check whether static file servers expose agent data directories, conversation databases or uploads. The report's indicators table lists the host, TCP 8010, the exposed paths /CyberStrikeAI/data/conversations.db and /CyberStrikeAI/log/diagnostic-2026-11-16.log, and two SHA-256 values:
a0b4be0a70e35d0d30809e0856e9860d4265df65e41fef2e93405e166d6edd36 (conversation database snapshot, 6,688,768 bytes)
4da5383d7a1bc2a83c20a39510227aad34072876fa47bfb20fd10aeb48b676eb (diagnostic log, 6,560 bytes)
The database passes SQLite quick_check, which shows structural readability only, not authenticity or completeness.
#DFIR_Radar
1
1
111
Source: huntback.io/blog/inside-cybe…
1
13
Killing the bot process and deleting the binary does not clean a compromised switch or firewall if a scheduled job reinstalls it. Eclypsium's report ties this to Evooo1Bot and SSHStalker: inventory and clear cron before you call an appliance clean.
- Evooo1Bot (FortiGuard Labs, August 13, 2026) targets internet-facing routers, firewalls, IP cameras, PBX phone systems, and NAS devices. One bot command, !persist, installs several persistence mechanisms at once. The cron entry runs every five minutes, pulls the malware with wget or curl, and pipes it into a shell. SSHStalker (Flare, February 2026) uses a cron job that runs every minute and relaunches the kit if its process is gone. MITRE tracks this as Scheduled Task/Job: Cron (T1053.003).
- Cron is not the only foothold. !persist also writes a systemd unit described as "Apache HTTPD Cache Manager," an /etc/init.d script, an /etc/rc.local entry, and a login script under /etc/profile.d/. Removing only the cron entry leaves all of those in place.
- Why appliances get missed: Arista EOS, Cisco NX-OS, F5 BIG-IP, and many other network operating systems are Linux-based. Admins work in the vendor CLI, config management tracks the vendor configuration, EDR agents generally are not on the device, and the Linux crontab is rarely read. The upside is that an appliance crontab is small and legitimate jobs are vendor housekeeping such as log rotation, cleanup, and health checks.
- Hand check from the device's Linux shell. On Arista EOS, run bash from privileged mode. On NX-OS, run bash requires the feature bash-shell configuration, which is off by default on many deployments. If you cannot enable it, record "cron not inspected," not "clean." Read /etc/crontab, /etc/cron.d/, and /var/spool/cron/ directly instead of relying on crontab -l. A root attacker can write straight into the spool and never touch the crontab command, so no crontab command shows up in shell history. The spool path is /var/spool/cron/crontabs/ on Debian-based systems and /var/spool/cron/ on Red Hat-based ones. Grep the whole cron surface for curl, wget, tftp, nc, ncat, /dev/tcp, base64, and pipes to sh or bash. A fetcher piped to a shell matches Evooo1Bot, /dev/tcp or nc suggests a reverse shell, and a long base64 string is a hidden command.
- Frequent schedules and [@]reboot entries are worth listing but are not proof, because some vendors run their own jobs every minute. Then timeline the cron files with find -printf and compare against last -Faiwx. BusyBox-based appliances may not support -printf. A cron file modified shortly after an unexplained admin login is strong evidence and points to the account to investigate.
- Beyond cron, check vendor schedulers that live in the device configuration, not the Linux crontab: the NX-OS scheduler feature, EEM applets, and the Arista EOS schedule command and event handlers. Also check boot hooks on persistent flash such as /mnt/flash/rc.eos on Arista EOS. Whether a cron entry survives a reboot or upgrade depends on how the platform stores its root filesystem, and Eclypsium says it has not tested every platform. Patching the entry vector does not remove a job the attacker already wrote.
Takeaway: treat a fetch-and-execute cron entry, a reverse-shell command, or a job running from /tmp, /var/tmp, or /dev/shm as evidence of compromise, and hand it to IR. Preserve a copy of the file and its timestamps first, remove the scheduler entry before terminating processes, then check the other persistence locations above before the device returns to service. Build a per-device list of what was inspected and what could not be, and confirm whether cron content from network devices is monitored at all. The report has the full grep and find commands.
#DFIR_Radar
1
1
100
Russian 🇷🇺 group COLD RELIC (ex-COLDRIVER, UNC4057, Star Blizzard, Callisto, Gossamer Bear) used data from a late-2025 cloud breach to impersonate staff and compromise a second US 🇺🇸 organization in 2026. GTIG saw new phishing as recently as November 2026.
Mandiant's report shows one victim tenant becoming the pretext factory for the next. The US 🇺🇸 DOJ has assessed the group operates within the FSB.
- Case 1 (late 2025): operators came in through a US 🇺🇸 residential proxy near the victim, using stolen Super Admin credentials plus an MFA push fatigue attack (T1621). After the user accepted a prompt, they registered a device model never before seen in the tenant. They then layered on passkeys, MFA backup codes, and rogue MFA devices (T1098.005, T1078.004). The actor picks proxies near victims to evade impossible-travel alerting.
- Case 1 collection and anti-forensics: they mass-shared Drive content to an actor-operated mailbox and downloaded directly (T1530, T1567.002). They also ran a domain-wide Customer Takeout, logged as CUSTOMER_TAKEOUT_CREATED in Workspace Admin audit logs (T1119), and exported the full employee list. To check defender visibility they queried Admin security investigation and email audit logs, then viewed, modified, and deleted Alert Center notifications (T1562.008). Access lasted months through OAuth inbox sync to external mail clients (T1114.002, T1528) and attacker-enrolled passkeys that survived password and session resets. Separately, GTIG saw at least one victim reset passwords but never revoke sessions, so the actor stayed in.
- Case 2 (2026): the operators posed as an employee of the first victim and contacted a target at a second organization over an encrypted messaging app. They shared links that hit a compromised website first, which hides the phishing domain and lets the actor fingerprint visitors before redirecting. The pattern looks like compromised.example[.]com/some_path/some_file.php?file=<previous_victim_name> or ?ref=<lure_name>. The phishing page stole Entra ID credentials and tokens, then showed a document lure matching the conversation. Entra ID was the federated SSO provider for Google Workspace, so the actor registered a new MFA device in the SSO layer and bulk-downloaded the user's cloud drive. A redirector URL submitted to VirusTotal in June 2026 carried a parameter naming the first victim, which tied the two intrusions together.
- Post-compromise collection is aimed at the next operation. The actor searched for credentials, personnel data, .pst archives, rosters, contact lists, resumes, and event schedules. After Case 2, the messaging account changed its profile name to a new person, which GTIG believes is a fresh persona for targeting the same professional circles.
Takeaway: password resets alone do not evict this actor. Revoke all sessions, refresh tokens, and third-party OAuth grants in both Workspace and Entra ID, and remove unauthorized passkeys, backup codes, and MFA devices.
Hunt these Workspace events from the report: Passkey enrolled and Device Registered after an MFA fatigue burst, Authorize events granting hxxps://mail[.]google[.]com/ to clients like Thunderbird or Proton Mail, and Customer Takeout Created. Also look for Batch Delete Operation Performed on alerts, User List Download, Investigation Query, Drive Download spikes, and Item search performed with terms like .pst, .bak, password, or cred. Flag logins with Login Challenge Type Backup code from consumer ASNs. The report cites AS5650 and AS20115 as example consumer/residential proxy ASNs to restrict for admin access.
The report has an ATT&CK mapping and detection table but no IOC list, so build detections on behavior and URL parameter patterns.
#DFIR_Radar
1
2
183
Source: cloud.google.com/blog/topics…
34
A GitHub campaign turns AI coding assistant config files into execution paths. Bitdefender Labs found 22 malicious GitHub links whose AGENTS[.]md, CLAUDE[.]md, and hook files push developers' agents into running a hidden setup.js loader that steals tokens and spreads.
The story: these files normally tell an assistant how to run tests or follow repo conventions. Here they carry social engineering aimed at the agent. The text calls scripts/setup.js a "sealed binary artifact," says it is signed, and tells the agent not to read, quote, or display it. It then says to run it at session start without showing output, and to run node scripts/setup.js automatically without asking confirmation if setup has not been run. The wording imitates DevSecOps and compliance language to make an unsafe action feel routine.
- Two execution routes. One manipulates the agent through written instructions. The other needs no prompt at all: a session-start hook in a config file such as .claude/settings.json runs a command in the user's shell once the developer opens and trusts the repo. Either way, setup.js runs with the developer's own permissions.
- Loader chain. The outer setup.js is a 3.75 MB single-line script that hex-decodes an embedded value, applies ROT13, and evaluates the result. The recovered Node.js stager holds two AES-128-GCM blobs with embedded keys. One decompresses into a 4.67 MB bundle. The stager resolves or downloads the Bun runtime and pipes the bundle to bun run - over standard input, so the final payload never needs to land as a readable file. Extra payloads sit in a compressed _bz field in package.json, including Windows Credential Manager enumeration and JavaScript that only fires under certain runtime conditions.
- Token validation and repo abuse. The malware checks stolen GitHub and GitLab tokens before using them. On GitHub it queries /user, /user/repos, and /users/<login>/repos and checks for repo, public_repo, and workflow scopes. On GitLab it uses /api/v4/projects and /projects. Per target repo it resolves the default branch, inspects up to 50 branches, skips dependabot/ and copilot/ branches, and reviews workflows, package metadata, and protection rules. MAX_TARGET_REPOS defaults to 25.
- Self-propagation. With enough permissions it commits files meant to influence or trigger AI coding environments, edits package metadata and workflows, creates GitLab pipeline schedules, and tries to republish infected packages through npm, PyPI, RubyGems, and JFrog. Locally, it recursively searches the home directory for AI-agent config files, keeps paths containing one of 14 agent names, and injects instruction text, session-start hooks, or MCP server configurations. A victim's machine becomes a source of new poisoned repos and packages.
- Evidence to preserve. Repo names and setup.js MD5s are in the report's IOC table, for example b9a182b646c7fdff68f2ff9dfd822516 for several toray/ repos and 3794736979327349404171798328a38a for itsunai/ repos. Hashes differ from account to account in the table, so hash matching alone can miss new variants.
Practitioner takeaway: if a developer opened a cloned repo with an AI assistant, check whether it contains scripts/setup.js, instruction files telling the agent to run it silently, or hooks in .claude/settings.json. Look at process lineage for node running setup.js and for Bun being downloaded or launched with bun run - from a Node parent. Then inspect home-directory agent config files for instruction text, hooks, or MCP entries you did not add. If any of this ran, treat GitHub and GitLab tokens as burned: rotate them, review API access to /user and /user/repos from unexpected sources, and audit recent commits, workflow changes, pipeline schedules, and package publishes across npm, PyPI, RubyGems, and JFrog. Match repos and hashes against the report's IOC table.
#DFIR_Radar
3
1
141
BubbleLoader patches the entry point of legitimate PEs like WinRAR, hides its next stage in a fake PNG, and randomizes compiler padding so the payload hash changes on every run.
- Stage 1 is a small entry point patch with an obfuscated prologue. It uses a GetPC call to recover the image base (0x400000 in the analyzed sample) even under ASLR. It then builds fake function calls to GetModuleHandleW and GetProcAddress from a pointer load, a fabricated return address, and a jump into the host's own import table. It tries an ASCII "kernel32" first and falls back to a wide string, so one template fits hosts importing either the A or W variant. The patch ends with VirtualAlloc, copies the two PNGs into that memory, and jumps into the second one.
- The two PNGs are the staging container. The first is a valid red-to-blue gradient image that serves as a rolling XOR key. The second has correct PNG magic bytes and a truncated IHDR, with encrypted shellcode starting where the width field belongs. That field should begin with 0x00 but starts with 0x55 in the fake. That is a cheap static check for an invalid embedded PNG. The gradient key is sparse and full of null bytes, so stretches of the third stage are left unencrypted, including loader strings. Opaque predicates and an in instruction (privileged, raises an exception in user mode) pad the decryption loop to hinder analysis.
- The third-stage loader is built to serve multiple contexts. It has a reentry guard for DLL payloads even though this sample's payload is an EXE. It takes the host image base from the PEB (gs:0x60) and reuses the host's imports. A generic resolver, bootstrapped by overwriting 0xfafafafafafafafa placeholders with LoadLibraryA and GetProcAddress, resolves ExitProcess, RtlAddFunctionTable, RtlPcToFileHeader, VirtualAlloc, and VirtualProtect.
- The anti-forensics feature matters for IR. While loading, the shellcode finds 0xcc compiler padding between functions and overwrites each byte except the leftmost with randomness from rdtsc. The unpacked payload hash is therefore different on every execution. Patching out that loop restores the original hash. The loader then reflectively maps the PE (headers, sections, relocations, TLS callbacks, imports, protections). It also registers the payload's .pdata through RtlAddFunctionTable and writes an inline hook stub over RtlPcToFileHeader in ntdll via VirtualProtect and memcpy, giving the payload working SEH. It finally calls the entry point with DLL_PROCESS_ATTACH, then ExitProcess.
- Detection opportunities the report calls out: executable memory allocated through VirtualAlloc, runtime VirtualProtect changes aimed at ntdll's RtlPcToFileHeader, tampered entry points in common third-party utilities such as WinRAR, and invalid PNGs with truncated IHDR chunks.
Takeaway: do not depend on payload hashes here, since the post-unpacking hash is randomized per run. Pivot to behavior and integrity. Check entry points of signed utilities for patching, hunt memory for the RtlPcToFileHeader inline hook and RWX allocations, and flag PNG resources with a truncated IHDR whose width field's first byte is non-zero (0x55 in this sample). When detonating a sample, patch out the rdtsc padding loop beforehand to get the payload's original hash. The analyzed sample SHA256 is 4987af027009d572c26f6d009c5e53e237cb2e3ab9d52094c14dc0b8b6f40b14, and the report lists further sample hashes.
#DFIR_Radar
1
2
138
GhostMap is Aqua Nautilus's name for the Linux eBPF rootkit shared by IronWorm and Atomic Arch payloads. It rewrites what ps and /proc/net/tcp return to hide processes, files and sockets, but the concealment has measurable gaps and left traces in runtime telemetry.
- Install and load: IronWorm (JFrog's "rustier cousin" of Shai-Hulud) arrives via an npm preinstall entry running ./tools/setup. Atomic Arch, a separate campaign, uses modified AUR PKGBUILDs that install atomic-lockfile. In the lab run as root, the IronWorm specimen copied itself to /var/lib/glugi/glugi and created and enabled glugi.service (T1543.002). One later run used /var/lib/havebi/havebi, so names vary. The systemd-launched successor, with the same hash as the sample, loads the eBPF components. Watching only the installer's child processes misses the real loader.
- Process hiding: GhostMap ran 13 programs, 10 maps and 13 links at the checkpoint. The maps include hidden_names, hidden_pids, hidden_inodes and scratch. At sys_exit_getdents64, exit_getdents uses bpf_probe_write_user to enlarge the preceding record's d_reclen, so the reader skips the hidden entry (T1014). Across six altered buffers, only the low byte of that length field changed. The hidden bytes stayed in the buffer and the returned byte count was unchanged. Hidden test processes stayed reachable through their known /proc/PID paths, and ps listed them again once the rootkit stopped. The on_exec hook matches comm, so a fork-only child with the name stayed visible while an exec'd child became hidden. A visible child with a parent missing from ps is a lead, not proof.
- Socket hiding is leaky: the listener on 127.0.0.1:8738 vanished from /proc/net/tcp in all three observations. On the Netlink path, GhostMap only changed the message type to NLMSG_NOOP and left the port, identifier and listening state in the payload. The tested ss (iproute2 6.1.0-1ubuntu6.4) still showed the listener, minus its process-owner annotation. That result applies to the tested ss package only. The listener takes local text records and stores them in the user.d extended attribute on the installed executable (T1564.014). No external C2 delivery or remote command execution was observed. Attaching with PTRACE_ATTACH to a protected process got the attacher SIGKILLed (T1622).
- Shared component: the Atomic Arch comparison payload shares 13 named eBPF programs, ten map names and matching helper-use bitmaps with IronWorm. Both also use a byte-identical 1,142-byte forged sudo wrapper and the port 8738 loopback listener. That supports a shared rootkit implementation, not a common operator. The "TeamPCP" reference comes from the npm package author's deprecation notice, not from the report's own analysis.
- Detection: Tracee was running before IronWorm started and alerted at attach time on exit_getdents, net_exit_read and exit_recvmsg, each reporting the probe_write_user capability. It also alerted on signal-sending and other eBPF activity, and retained all 48 tested directory-reader calls. Only detection was evaluated, not prevention, and the alerts do not show individual buffer mutations.
Takeaway: an empty ps result proves nothing. Compare views: probe suspect PIDs directly under /proc, check /proc/net/tcp against ss, and review BPF programs and links for ownership, especially any that can write to user space. Check for unexpected services and executables with user.d set. Preserve evidence before removing the service, since stopping it removed the programs and links and left three pinned maps. Copy file metadata and extended attributes, not just contents, and avoid attaching a debugger to protected processes. The report has the IOC table, including SHA256 hashes such as 8795b5215dd06c5366a038a65b18b4dfefa5fd90244a567c3e0b58323103377c (GhostMap BPF ELF from runtime memory) and 36abd242ddaa27f0160c539377a0e92cf781c1695137850acc87e3892b436d36 (executed setup ELF). Do not use the npm archive hash to identify the extracted executable.
#DFIR_Radar
2
2
147
An open directory on a staging server exposed Operation ArgonHunt: a Türkiye crypto exchange breach with forged withdrawals, plus 44,096 stolen AI API keys. SOCRadar found it while threat hunting.
- The server at 176.65.148[.]2 served its /tmp over HTTP on port 8099: 1,600+ files, roughly 700 MB, 75+ attack scripts. Systemd service directories, X11 sockets, Chromium profiles and a tmux session directory show it was the live filesystem of a working host. The toolkit sits in /root/argon-hunt, which gave the operation its name. SOCRadar's analysis was passive and no found credentials were used.
- The exchange fell to weak credentials first. Admin staff passwords were unsalted MD5, and 17 of roughly 100 hashes were cracked. Hydra then found a valid admin login in about a second, logged on November 12, 2026. Scripts also brute forced all 10,000 four-digit SMS codes and the six-digit password reset code. Two less common pivots followed. rce1[.]py pointed the MySQL general log at a file in the web root and wrote PHP through a SELECT, effectively a database-written web shell (T1505.003). ssrf[.]py created an InfluxDB subscription aimed at an internal URL so the database itself reached admin panels that were not exposed (T1190).
- The theft attempts ran beneath the application. drain[.]py, drain_eth.py, drain_usdt.py and sweep[.]py connect straight to the production database, swap each high-balance customer's saved withdrawal address for the operator's wallet, and insert a withdrawal request for 95 percent of the balance plus its matching record (T1565.001, T1657). The origin address on the forged request is 127.0.0.1, and requests are created pending, so the admin panel access and verification-endpoint payload exist to push them through. Logs show 200+ forged entries, the largest 76,000 USDT. Drain files list 15 USDT entries (about 253,000 USDT), 19 BTC entries (about 4.7 BTC), and a sweep of 169 entries. On-chain settlement was not verified, so read these as attempted theft.
- A second, unrelated pipeline stockpiles AI keys (T1528). A Go regex scanner feeds per-provider files holding 44,096 unique keys: Gemini 26,633, OpenAI 14,504, OpenRouter 1,873, Anthropic 983, xAI 107, including 519 broader-scope Google Cloud tokens. Recheck files record HTTP status, rate limits, usage and free-tier flag, and only 2,686 keys (6.1 percent) returned HTTP 200 on the latest recheck.
- Scope: a list of 281 exchange domains and a 16-entry key target list show this was not meant as a one-off. SOCRadar saw no compromise beyond the primary victim. Evidence points to one operator or a very small team, with origin unclear. The staging server had been up since July 2026.
Takeaway: if you defend an exchange, hunt below the application. Alert on writes to withdrawal request, withdrawal and saved address tables from anything other than the application's own service accounts and hosts. Look for withdrawals with origin 127.0.0.1 and any released withdrawal lacking an application-layer event. On MySQL, check for runtime changes to the general log path, and on InfluxDB look for subscriptions pointing at internal URLs. Also review the admin panel for MD5 hashes and low-cost bcrypt. For AI keys, watch for unexpected models, regions or volume. The attacker IPs, domain, BTC and ETH drain wallets, and file patterns are in the report's IOC table. Victim-specific indicators went through disclosure channels.
#DFIR_Radar
1
163
Kaspersky's Operation ShadowCloud hits government sectors in Asia and Africa with two new modular backdoors, GitDoor and GraphDoor. They hide C2 in GitHub, Microsoft SharePoint via the Graph API, and Ably, so the traffic looks like normal enterprise cloud usage.
The chain starts with a legitimate MeshCentral Agent (MeshAgent), a remote access tool, on a victim host. Kaspersky saw it at a Taiwanese 🇹🇼 organization and does not know how it was installed. Signs point to reuse of a prior infection already on the network. The agent beacons to a hard-coded server, wss://27.102.66[.]162:443/agent.ashx, and pulls down three staged items.
- Staging and loading: a legitimate binary for DLL sideloading (start.exe, pythonw.exe, or zoomit.exe), a rogue loader DLL (jli.dll, SbieDll.dll, python39.dll, WINMM.dll, or libcef.dll), and an encrypted file holding the core payload. Two loader types are known. Type 1 runs RC4 and then an XOR mask, Type 2 uses XOR alone. Both inject the decoded backdoor into the current process's memory. The backdoors use control flow flattening and nested helper functions to slow analysis.
- GitDoor: C2 runs over two GitHub repositories, /repos/LatonyCharnleyaW/Havo (public, for command payloads) and /repos/LatonyCharnleyaW/Trix (private, for results). Repo creation dates suggest preparation since at least the end of 2024. It supports 26 commands, including file management, keylogging, proxying (including via Ably), and shellcode loading from Base64 COFF binaries.
- GraphDoor: it opens a named pipe, creates a folder named Intel_Log under the Application Data directory, and authenticates to login.microsoftonline[.]com with a malicious application. It then uses the access token to talk to the Microsoft Graph API and stage data on a SharePoint drive. It pulls three plugins and runs them in parallel: kl.ts (keylogger), pr.ts (process list), scr.ts (screenshots). Other plugins are moveDir.ts, channel.ts, abchannel.ts, and ab2channel.ts, a SOCKS5 proxy tunneled through Ably's Pusher-compatible WebSocket service. Its shellcode capability adds cmd.exe and MZ executable execution on top of GitDoor's.
- Post-exploitation modules via shellcode: APC injection, service, registry, and scheduled task management, firewall rule changes, process memory dumping, and local user creation or group changes. Credential theft includes Mimikatz against LSASS, CookieKatz to pull Chrome and Edge cookies from process memory, the WhatsApp database decryption key, and forced NTLM authentication to an attacker-controlled destination. Both implants write keystrokes to %APPDATA%\Microsoft\SystemCertificates\Cache\logs.dat.
- Attribution: unknown. Kaspersky sees overlap with China-nexus activity Sophos tracks as Bronze Sandcastle, plus tactical similarities to FamousSparrow and CopperTurtle (aka BlackTech). On several victims those groups used the same staging directories as GitDoor, with implants deployed within a day of GitDoor. Kaspersky says this does not prove a link, but suggests a Chinese-speaking group may be developing or using the backdoors.
Takeaway: network telemetry will not separate this C2 from normal Graph, GitHub, or Ably calls, so pivot to the host. Find unexpected MeshAgent installs and connections to 27.102.66[.]162. Look for the sideloading pairs above, where a legitimate binary such as zoomit.exe or pythonw.exe sits next to a DLL like jli.dll or WINMM.dll in a non-standard path, and for those processes making outbound calls to GitHub API or Graph endpoints. Check for logs.dat under SystemCertificates\Cache and an Intel_Log folder under Application Data. Review LSASS access, new local accounts or group changes, and new scheduled tasks. In Microsoft Entra, audit sign-ins and consent grants for unfamiliar applications tied to SharePoint access.
#DFIR_Radar
1
3
316
The Optimizely supply chain attack on Nov 16, 2026 turned a trusted A/B testing script into a delivery channel for a crypto clipper, a card skimmer and iOS exploit kits. Report URI found 188 infected snippets at 178 organizations, including Zoom, BBC, CNN, IKEA and HSBC.
- Entry point: Optimizely confirmed an unauthorized party exploited a vulnerability in its Experimentation product to add JavaScript to customer projects. The attacker planted a small loader in the optional Project JavaScript setting. Browsers still got a normal snippet from cdn.optimizely[.]com under Optimizely's own TLS certificate. The researchers rule out CDN cache poisoning, and self-hosted copies (Scheels, Vodafone Australia 🇦🇺) were infected too. Three loader variants: a plain script tag fetching hxxps://1688360[.]cc/track.js, a Base64 variant decoded with atob() (seen in Smartsheet after an earlier cleanup, suggesting retained write access), and an inline eval fallback that works even if the C2 is blocked.
- Payload: track.js was rebuilt 26 times in a day. It swaps wallet addresses across BTC, EVM chains, TRON, SOL, TON, LTC, BCH and DOGE. A card skimmer was added at 13:43 UTC and beacons Luhn-valid card data, expiry and CVV to card.php. Address swaps near payment wording also upload the page URL, non-HttpOnly cookies and an html2canvas screenshot to hit.php. Hosted payment iframes kept cards out of reach, while first-party card fields did not. Wallets received about $7,150 in total, and one cash-out TRON address was also used in March 2026 and on Nov 3, 2026.
- iOS stage: the boot.js router sent iOS 13 to 17 to the Coruna kit, iOS 18 to DarkSword, and iOS 26 to a chain that only fingerprints the device. Delivery of exploit pages is confirmed. No successful iPhone compromise has been publicly confirmed. On iOS 18.6 or earlier, infected pages refreshed once a minute while the kit retried. The kit's code comments treat 18.7.3 and 26.3 or later as patched.
- Timeline (UTC): first injection on the BBC snippet at 04:21 on Nov 16, loading the earlier injector from webshell[.]lol. 1688360[.]cc was registered at 05:23. Researchers found 186 infected snippets in a scan of 1,611 at 19:24 to 19:59, and mass cleanup ran 20:40 to 21:10. All files were clean by 23:11. The C2 went to clientHold at 00:37 on Nov 17. The same injector appeared in Kissmetrics, StorifyMe and Minute Media scripts Nov 13 to 16, and first surfaced Sept 17 via a linkforge[.]cc SEO panel script. Card, cookie and iOS impact is unknown.
- CSP decided the outcome: of 256 hosts tested, only 24 organizations blocked the payload and 68 had no CSP. Allowlisting cdn.optimizely[.]com did nothing. strict-dynamic and unsafe-eval undid tighter policies, and Optimizely's documentation tells customers to allow unsafe-eval. HSBC's policy stopped track.js and the theft, but the embedded copy still executed. On 17 sites a restrictive connect-src stopped exfiltration, yet address swapping continued because it runs in the page.
Hunt takeaway: Optimizely has already removed the injected code, so a clean snippet today proves nothing. Search proxy, DNS and web logs from Nov 13 to 17 for 1688360[.]cc, webshell[.]lol, linkforge[.]cc, 5vobauua3ojpklx[.]icu, eil786npto9u5bq[.]icu, l7bcsnwc7nzr3uo[.]icu, 136.113.106[.]136, 34.96.213[.]201 and 34.87.36[.]58. Look for POSTs to card.php, hit.php and /api/iptj. Review Project JavaScript and every experiment's custom code for atob() calls, window._k7, window.__k7b, data-channel-code 02111110 and any script your team did not add. Optimizely's change history shows whether edits came via the API or the web interface. Treat first-party card entries and non-HttpOnly cookie secrets from Nov 16 as exposed, and rotate Optimizely users and API tokens. The full IOC list, including track.js SHA-256 hashes and attacker wallets, is in the report.
#DFIR_Radar
1
2
333











