Saudi aviation and petrochemical, with a wiper in the toolkit.
Also tracked as APT 33 · Elfin · MAGNALLIUM · Refined Kitten · HOLMIUM · COBALT TRINITY · G0064 · ATK35 · Peach Sandstorm · TA451 · BLEAK ION · REFINED KITTEN
Summary
APT33 is a threat actor attributed to Iran (Islamic Republic of), assessed as espionage-motivated. Its recorded targeting runs to private sector across 3 countries in the open reporting we hold — Saudi Arabia among them.
Across 31 attributed ATT&CK techniques we resolve 12 distinct behaviours — the things they do that a detection can be written against. The two that account for most of their observed tradecraft are valid account abuse or account creation and persistence pointing into world-writable staging. Their toolkit is 16 named items: 5 we see only in this actor's reporting and 11 commodity or dual-use. That ratio is the point — the first group is what identifies them, the second is what they share with every other group and with red teams.
For a Gulf defender the relevant line is the victimology: Saudi Arabia appear in the reporting behind this record. That is evidence of past targeting, not a forecast, and it is drawn from what vendors chose to publish — an actor working quietly in a market nobody reports on looks identical to one that is absent.
As MITRE describes them: “Our analysis reveals that APT33 is a capable group that has carried out cyber espionage operations since at least 2013. We assess APT33 works at the behest of the Iranian government.” — MITRE ATT&CK, CC BY 4.0.
On the name. “APT33” labels a cluster of related activity tracked under one name — not a verified organisation. Whether one team, several contractors or a rotating cast sits behind it is not something this record establishes, and the vendor names below are separate groupings that only mostly overlap.
02 — The detection plan
Most actor profiles hand you indicators that expire in days. These are the behaviours the actor cannot drop without changing how they operate, each with the reason it holds. Weight is our own: how much of the actor's observed tradecraft the behaviour accounts for.
The persistence mechanism and the directory it points at both outlive the payload. The discriminator is the TARGET PATH, not the task or key name, which the actor can change freely.
8 techniques · weight 5
Few legitimate processes read LSASS memory, so precision is high in most estates once the agents that do are excluded.
2 techniques · weight 5
The delivery wrapper changes constantly; an office, archive or browser process parenting a shell is the same shape across all of them.
8 techniques · weight 5
Kept separate from the LSASS rule because it needs different telemetry and tunes differently — a browser credential store read by a non-browser process is its own signal.
8 techniques · weight 4
The command-line SHAPE — the flag combination, not the payload — is stable for months to years and cheap to match.
8 techniques · weight 4
Visible only in identity telemetry, and the failure-then-success shape is what distinguishes it from ordinary user error.
6 techniques · weight 3
The set of system binaries that can fetch a file is small and changes with the OS, not with the actor.
5 techniques · weight 3
Staging is a step the actor cannot skip, and it happens before the data leaves — the last cheap place to catch it.
7 techniques · weight 3
Not a durable behavioural signature but the entry vector for a large public estate — worth an exposure control even where the detection is weak.
3 techniques · weight 3
The tunneller is a tool choice that persists for months, and the traffic shape is visible even when the payload is not.
8 techniques · weight 3
Produces no malware event at all — the only trace is in identity telemetry, which is why an endpoint-only programme has a blind spot exactly here.
10 techniques · weight 3
The URI shape and cadence are cheap to match and hold for weeks to months — but the destination itself rotates in days, so match the pattern rather than the host.
7 techniques · weight 3
No single behaviour here is worth waking someone. Summed weights of 6 on one host inside a window is an alert; 10 is an incident. The pair that most reliably means this actor rather than commodity crime is Persistence pointing into world-writable staging + Credential access against LSASS — seen together, treat as an incident regardless of total.
03 — Tooling
Purpose is derived, not asserted: each tool's own ATT&CK techniques are resolved to tactics, and the tactics it spends most of its techniques in are what it is for.
Actor-specific means we have no record of it outside this actor's reporting — those are the names that identify them. Commodity covers everything sold, published or shared: Cobalt Strike and Brute Ratel are licensed red-team software every pentester owns, Mimikatz and Impacket are public, and backdoors like ShadowPad circulate between unrelated groups. Finding a commodity tool tells you far less, which is exactly why the split is drawn. ATT&CK's own “malware vs tool” type answers a different question — whether the software is malicious — so it is not used for this.
Actor-specific — 5 items
Destruction or extortion and gathering data.
Destruction or extortion and evading defences.
Credential theft and staying resident.
Gathering data and staying resident.
Privilege escalation and remote control.
Commodity and dual-use — 11 items
Credential theft and gathering data.
Gathering data and staying resident.
Credential theft and gathering data.
Credential theft and gathering data.
Credential theft and gathering data.
Credential theft and moving between hosts.
Staying resident and moving between hosts.
Gathering data and staying resident.
Staying resident and network reconnaissance.
Moving data out and moving between hosts.
File hashes — actor-specific families — 5 held
| Hash | Type | Family | First seen |
|---|---|---|---|
| 0b1de625a89da12bd1fdd292b341bad3 | md5 | AutoIt backdoor | 2026-07-24 |
| 07ed2c9ed61b60078af0164f061696be | md5 | AutoIt backdoor | 2026-07-24 |
| 07bb21d28ae4ab07d62f8deb4343aaeb | md5 | AutoIt backdoor | 2026-07-24 |
| 05c07339603994b36dcfefcce720d03d | md5 | AutoIt backdoor | 2026-07-24 |
| 03e4bef86f3e3e6ea23eb6f017af0c98 | md5 | AutoIt backdoor | 2026-07-24 |
04 — Coverage, honestly counted
Of the techniques attributed to APT33, 27 fold into the detection plan above, 1 describe preparation you cannot see from inside your network, and 3 are uncatalogued — we have not yet placed them. That last number is published on purpose; it is the honest size of the gap.
05 — The chain
The same techniques as above, sequenced. Useful for deciding where to spend a detection: the earlier a tactic sits, the more of the intrusion you still get to prevent.
06 — Tradecraft artefacts
Pulled from primary vendor research and kept attributed. These are hunt starting points, not detections — a path an actor used once is worth a query, not a rule.
| Artefact | Kind | First reported by | Date |
|---|---|---|---|
| %programdata%\info\info.txt | file path | Recorded Future | 2026-02-19 |
07 — Sourcing, graded
NATO Admiralty grading. The letter is the source's reliability, the number the credibility of the claim. We separate primary vendor research from outlets reporting on it, because ten articles derived from one report is one report.
| Grade | Source | Basis |
|---|---|---|
| B2 | MITRE actor record | 19 references across 17 distinct research hosts |
| C3 | Cloudsek | 7 items — reporting on other vendors' research |
| C3 | The Hacker News | 4 items — reporting on other vendors' research |
| B2 | Palo Alto Networks Unit 42 | 2 items — primary vendor research |
| B2 | Recorded Future | 2 items — primary vendor research |
| B2 | Group-IB | 1 item — primary vendor research |
| — | Estate telemetry | no data — absent, not negative |
16 items · 5 outlets · 44% of reporting traces to a single outlet
08 — What we do not know
Every vendor profile has these. Most omit them, which leaves you unable to tell a quiet actor from an unobserved one.
No URI paths, headers or beacon sequences for any family we hold, and no URL indicators at all to derive them from — the indicator set is hashes, hosts and addresses.
No domains are attributed to this actor in our holdings, so no naming convention, TLD preference or hosting pattern can be derived.
09 — Outlook
Probability language is ICD 203. Likelihood and confidence are stated separately: how likely we think it is, and how good our basis is for thinking so.
11 — Related actors
Ranked by shared victimology first, shared attribution second. The reason is on each card.
Credential phishing at scale against Gulf government and academic mail.
Travel, telecom and personal-movement data across Kuwait and Saudi Arabia.
Maritime and research targets, Saudi Arabia included.
10 — Primary sources
19 of them. Published in full so the page can be checked rather than believed.
Run it against your own estate
The library is the part we can publish. The platform runs the same detection plan against your own estate, tells you which of these behaviours you already cover, and retro-hunts the rest — on your hardware, with nothing leaving the building.