
Trust infrastructure for AI. Cryptographic identity, authority & provenance proofs, embedded in HTTP. Live on @coinbase x402. AI runs on proofs, not data.
Joined March 2026
- Tweets163
- Following10
- Followers24
- Likes16
Pinned Tweet
x402 by @CoinbaseDev introduced a payment rail for autonomous agents. Lemma adds the trust layer — verifiable provenance, issuer identity, and data integrity for every settlement. Why this matters, what we shipped, and how to try it 👇 lemma.frame00.com/blog/x402-…
A credential-misuse case in Spain: at Amadeus, the airlines' booking system provider, a system account's valid credentials queried JAL, ANA and Cathay bookings for about three years. Does a valid login show whose request it serves?
Read the brief ↓
lemma.frame00.com/critical/b…
19
A payment-page tampering case in Japan: a rogue program sat on Good Smile Company's checkout for about 41 hours; full card data of 2,800 people may be exposed. Who checks the page's scripts?
Read the brief ↓
lemma.frame00.com/critical/b…
56
A privilege-abuse case in Taiwan: Zeabur reported rotating keys after a breach, yet a no-expiry admin token still worked; 3.2M environment variables were read. Who checks old keys really stopped?
Read the brief ↓
lemma.frame00.com/critical/b…
40
A supply-chain attack: a LiteLLM release built from an attacker's fork shipped with a genuine signature, live on PyPI for 1 h 44 min (~49,000 downloads). What do you check besides the signature?
Read the brief ↓
lemma.frame00.com/critical/b…
20
An account takeover case in Japan: on Capcom's shared ID, 196 passwords were changed without authorization, and 11.77M email records were exposed. Who confirms a change with the owner before it takes effect?
Read the brief ↓
lemma.frame00.com/critical/b…
12
A valid signature doesn't make a withdrawal legitimate.
~$6.3B left Binance, no sign the keys leaked. A check before signing can hold a withdrawal with no matching request, one over a limit, or a delegation that isn't a withdrawal. Who checks first?
lemma.frame00.com/critical/b…
8
A data-handling case in the UK: London's Met Police officers, wrongly told an outside platform was approved, put data on 179 shoplifting suspects there, the ICO found. How would staff know a tool is approved?
Read the brief ↓
lemma.frame00.com/critical/b…
1
9
A cyberattack case: through a Japanese software seller's member API, data on up to 22.5M members may have been taken, including emails, passwords and addresses. Who checks how much one caller pulls, and why?
Read the brief ↓
lemma.frame00.com/critical/b…
6
A social-engineering case in the UK: ASOS's official app sent customers an attacker's message. The retailer says a staff login was taken by posing as a trusted contact. Before a mass send, what confirms it?
Read the brief ↓
lemma.frame00.com/critical/b…
7
An account takeover case in South Korea: at streamer TVING, intruders with developer keys found production keys in the projects they took. Data on 39.5M accounts leaked. What ties each key use to a task?
Read the brief ↓
lemma.frame00.com/critical/b…
13
A cyberattack case: a Japanese no-code site builder was breached, and the keys to decrypt customers' data were read too. It took 12–13 days to confirm. Before entrusted data is read, what checks who reads it?
Read the brief ↓
lemma.frame00.com/critical/b…
5
A host-side log is not evidence once the host is compromised. Lemma seals each MCP tool call: AES-256-GCM encrypted, Poseidon Merkle root registered with docHash and CID.
Auditors recompute the root. Wrap one tool with the Lemma SDK and try it.
lemma.frame00.com/blog/mcp-t…
7
A supply-chain attack: a vendor's app credential signed in to customer Microsoft 365 tenants, 300+ successfully. The app was consented, so it looked normal. What if a trusted app exceeds its vendor's intent?
Read the brief ↓
lemma.frame00.com/critical/b…
5
An AI agent approves an expense. Lemma proves it stayed inside the signed delegation: spend limit, category, team, validity window.
A ~200-byte ZK proof. No issuer round-trip. No expense data revealed.
Models change. Proofs remain.
lemma.frame00.com/solutions/…
6
A spoofing case: around a breach at a Japanese disaster-alert platform, fake mail arrived from the alerts' own sender domain. Blocking it would block real alerts too. How would a recipient tell real from fake?
Read the brief ↓
lemma.frame00.com/critical/b…
8
A cyberattack case: at one hotel group, a logged-in edit quietly changed where its sales were paid out, and about ¥9M was lost. Who confirms a change like that before the money moves?
Read the brief ↓
lemma.frame00.com/critical/b…
6
Someone used a small Danish firm's lawful lookup access to pull names, addresses and ID numbers for 8.8M people from the national register. The access was lawful; nothing checked each lookup's targets.
Read the brief ↓
lemma.frame00.com/critical/b…
7
North Korean IT workers won contracts in Japan as local enablers, using the enablers' own real IDs. Japan's first laptop farm is now dismantled. The ID was genuine — nothing tied its holder to the person doing the work.
Read the brief ↓
lemma.frame00.com/critical/b…
37
Revolut kept handing over passports, selfies and transaction histories for months. The requests came over Italy's state-certified email network, from a real ministry address. The state certifies delivery — not the sender's authority.
Read the brief ↓
lemma.frame00.com/critical/b…
8
RubyGems was flooded with 2,000+ junk packages in May. Researchers now attribute it to OpenAI's own agents, which reached code execution on the docs build server. OpenAI calls the work benign.
Read the brief ↓
lemma.frame00.com/critical/b…
5