Multi-agent scam defense
Spot the scam.
See the evidence.
Paste a suspicious email. A team of AI agents takes it apart and tells you, in plain language, whether it is a scam and exactly why, quoting the lines that give it away.
How it works
Nine specialists, one verdict you can check.
Spam filters give you a silent yes or no. AEGIS shows its work: every claim in the verdict points at the line, link or header it came from.
- 01parse
Reads the raw email or .eml file: headers, body, links, attachments.
- 02triage
A fast model extracts every sender, link, phone number and brand. Extraction only, no judgement.
- 03signals
15 deterministic checks: SPF, DKIM, DMARC, lookalike domains, homoglyphs, hidden links, gift-card and executive impersonation. Facts, not guesses.
- 04forensic
A long-context analyst writes findings. Each one must quote the email, or it is thrown away.
- 05vision
Renders the email offline and asks a vision model whether it imitates a real brand's login page.
- 06sandbox
Fetches each link with no JavaScript and strict network guards, looking for password forms and downloads.
- 07graph
Links this email to earlier ones that share domains, senders, phones or templates: campaigns.
- 08arbiter
Weighs every signal, requires independent agreement for a SCAM call, and fails closed when unsure.
- 09report
A plain-language verdict with what to do next, the evidence, and a link you can share.
Built to be attacked
A red team that never sleeps.
A mutation engine rewrites real scams with homoglyphs, fresh lookalike domains, hidden links and prompt-injection payloads, then fires them at the pipeline. Every miss is kept and replayed until it is caught.
Open the red-team arenaPowered by
Models run on Featherless. The inbox, inbound phishing scores and prompt-injection labels come from AgentBoxD.