Forager
Reads inbound mail and malware behaviour for the regional attacks generic filters miss.
Localised, multilingual scams that generic filters miss get caught and explained.
Inbound mail is analysed for impersonation, social engineering, suspicious links and attachments, QR phishing and conversation hijacking, alongside mail-authentication, reputation and sandbox signals. Sandbox behaviour, imports, strings and decompiler output are summarised into candidate indicators and draft detection rules.
- Inbound email, including multilingual and localised scam patterns
- Sandbox behaviour reports, imports, strings and decompiler output
- Signed offline threat-intelligence packs, correlated against internal events
The mechanism, not the marketing. If a step below would not survive an engineer's scrutiny, it should not be on this page.
Operating path
From signal to auditable evidence
01Observed inputs
- Inbound email, including multilingual and localised scam patterns
- Sandbox behaviour reports, imports, strings and decompiler output
- Signed offline threat-intelligence packs, correlated against internal events
02Deterministic workflow
- 1 Authoritative signals come first SPF, DKIM and DMARC results, reputation data, and sandbox detonation output. The artifact is detonated in the sandbox — never reasoned about in the abstract, and never executed anywhere near the model runtime.
- 2 Language and region are first-class, not an afterthought Localised social engineering in Bahasa, Vietnamese, Thai and code-switched English is evaluated by a model that has those languages. An English-trained filter reads unfamiliar text as ordinary business correspondence, which is precisely the gap regional attackers use.
- 3 QR codes and image-borne links are extracted before evaluation When the payload is not in the text, analysing the text finds nothing. Extraction happens first, in the sandbox.
- 4 Thread continuity is checked against actual history A message claiming to continue a conversation is compared against the mail history that would have to exist for that claim to be true. Hijack and fabricated-thread attempts fail this check.
- 5 Threat intelligence arrives as signed offline packs Indicators are correlated against your own event store and carry provenance and expiry. Vendor intelligence, internal observation and model inference are never merged into a single confidence number.
- 6 Detection content is drafted, reviewed, then deployed Candidate indicators and hunting rules are produced for a human to approve. Nothing reaches your detection stack automatically.
03Evidence outputs
- Analysis records with mail-authentication, reputation and sandbox signals attached
- Candidate indicators with provenance and expiry
- Draft detection and hunting content with its reviewer disposition
The Bahasa invoice the English filter waved through
Authentication passed. No malicious link. No attachment. Scored clean by every conventional control.
The payment did not go out. Every conventional control had scored this clean, and each of them was working correctly: the attack was carried in language and in conversational context, not in a link or a file.
The authority ladder. Green rungs are authoritative — a rule, a scanner, a sandbox, or a person. Blue is where the model contributes. Amber is where a human decides.
What the model adds
Language and regional judgement — the Bahasa-Indonesian invoice scam or the Thai-language impersonation that an English-trained filter reads as ordinary business mail.
What it cannot claim
It cannot determine that an artifact is malicious on its own; the sandbox and the analyst do that.
- Never executes an artifact inside the model environment. Detonation happens in the sandbox.
- Never conflates vendor intelligence, internal observation and model inference — provenance is tracked and shown.
- Never lets an indicator outlive its expiry silently.
- Never deploys detection content without review.
- Analysis records with mail-authentication, reputation and sandbox signals attached
- Candidate indicators with provenance and expiry
- Draft detection and hunting content with its reviewer disposition
A specification, not a simulation
We will not show you a scripted demonstration of something that does not exist. What we publish instead is the workflow specification, held to the same standard as shipped work: actor, trigger, happy path, exceptions, terminal state, and the audit evidence it must leave. When it is built, this section becomes a demonstration.
Status and what promotion requires
Roadmap discovery. Promotion requires a named mail platform and sandbox, plus per-language detection-quality measurement.