What we do with your data, and how we defend it.
Frameworks we follow
- OWASP Top 102021
- Web applications: this site, its forms and its administration.
- OWASP API Security Top 102023
- Internal routes and the demonstration endpoint.
- OWASP Top 10 for LLM Applications2025
- The live demonstration and the monitoring pipeline behind the Signals: prompt injection, insecure output handling, sensitive information disclosure, unbounded consumption.
- OWASP Top 10 for Agentic Applications2025 (first edition)
- The agent systems we deliver to clients: goal hijacking, tool misuse, identity and privilege abuse.
- MITRE ATLAS2025
- Adversarial threats against the models we use, for the demonstration and for client systems.
Threat model of this site
Five surfaces, each with the threats we considered and the controls in place. Residual risk is reviewed before every major release and after any incident.
Contact form
Threats
Spam and automated submissions, injection through the fields, harvesting of leaders' contact details, cross-site request forgery.
Controls in place
- Five fields only, validated and typed server-side
- parameterised queries
- rate limits per address and per day
- CSRF protection
- explicit consent recorded with a timestamp
- data stored in the European Union
- access restricted to the house
Live demonstration
Threats
Direct prompt injection, attempts to extract instructions or data, abusive consumption, hostile or off-topic input, personal data pasted by the visitor.
Controls in place
- Input is untrusted data: bounded length, URLs refused before any model call, injection filter, strict separation between instructions and content
- the model has no tool that writes, sends or pays
- output validated against a schema before display
- rate limit per address and per day
- token budget per call and monthly cap with graceful degradation
- AI disclosure on the first line
- nothing retained without consent
Administration
Threats
Credential theft, session hijacking, privilege escalation, injection through content awaiting validation.
Controls in place
- Strong authentication with a second factor
- authorisation checked server-side on every route
- audit log of every publishing action
- content awaiting validation rendered as inert text
- strict Content Security Policy, HSTS, frame-ancestors denied
Monitoring pipeline (Signals)
Threats
Indirect prompt injection through a page the pipeline reads, publication of false or manipulated content, supply-chain compromise.
Controls in place
- Fetched content is untrusted and filtered before the model
- outbound requests limited to an allow-list of hosts
- nothing is published without human approval
- every Signal carries an AI disclosure
- dependencies tracked with a software bill of materials and continuous analysis
Outgoing e-mail
Threats
Spoofing of our address, unsolicited sending, leakage of a visitor's details to the wrong recipient.
Controls in place
- SPF, DKIM and DMARC enforced on the sending domain
- acknowledgements sent only to the address that submitted the form
- no commercial sending without prior consent, in line with Section 30A of the Israeli Communications Law and with Article L.34-5 of the French CPCE
What we do with your data
Minimisation
The form asks for five things: name, e-mail, company, country, subject. The demonstration asks for a description of your business, not for a name. We do not ask for more than we use.
Hosting in the European Union
The database and the server functions run in the European Union. Data subject to GDPR does not leave it, except for the transfers described below.
Transfers to Israel
People of the house may access data from Israel. That transfer is covered by the European Commission's adequacy decision for Israel, Decision 2011/61/EU, reviewed and confirmed on 15 January 2024.
Retention and erasure
Contact requests are kept for twenty-four months after the last exchange, then deleted. Technical logs of the demonstration are kept for thirty days, with the address hashed. Any request for erasure sent to hello@elahadvisory.ai is honoured within thirty days.
Processors
Vercel for hosting and functions; Neon for the database; Anthropic for the model behind the demonstration and the Signals, under a data processing agreement and without retention for training; a transactional e-mail provider for acknowledgements. Audience measurement, without cookies, is done by the site itself. Each is bound by a contract and listed in the privacy policy.
Compliance by country
- France
- GDPR: consent as the legal basis for the form, information at the point of collection, rights exercised by e-mail, supervisory authority the CNIL. Direct marketing by e-mail follows Article L.34-5 of the CPCE: no sending to an individual without prior consent, an opt-out in every message.
- Spain
- GDPR and the LOPDGDD: consent recorded, information in Spanish, rights exercised by e-mail, supervisory authority the AEPD. Legal notice under Article 10 of the LSSI; commercial e-mail under Article 21 of the LSSI, with prior consent.
- Israel
- Protection of Privacy Law 5741-1981 as amended by Amendment 13, in force since 14 August 2025, and the Data Security Regulations 5777-2017: documented purposes, access limited to the house, incident notification where required. Commercial messages follow Section 30A of the Communications Law: prior written consent, clear labelling, a right to refuse in every message.
Reporting a vulnerability
If you find a weakness in this site or in a system we run, write to us with the details and, if you can, a way to reproduce it. We acknowledge within seventy-two hours, we keep you informed until it is fixed, and we do not take action against good-faith research. A security.txt file at the root of the site carries the same contact.
subject line: Security