The Patch Anza Had to Ship Blind
A hash went public before the fix did.
Hey, it’s Arsen.
In today’s menu:
• Why Anza had to patch a validator bug without explaining it
• Why a “47% drop” in hack losses is actually a 28% rise
• Why AI agents already move billions in fractions of a cent
• And more…
🏴☠️ Vulnerability
A Validator Crash Bug Anza Had to Patch in Secret
You’re reviewing how Solana’s runtime loads a compiled program.
Every program ships as an ELF file. Real instructions live inside its .text section.
CALL_REG is a VM opcode that jumps to an address at runtime — but only works if that address lands on an instruction boundary.
Programs built through the normal Solana toolchain always align clean. The linker guarantees it.
The runtime also accepts programs built outside that toolchain.
What checks the alignment on those?
Nothing did.
// ELF sanitization checked structure, symbols, sections —
// never .text alignment
Deploy a hand-crafted ELF with a misaligned .text section, call it once, and CALL_REG jumps to a garbage address. The host process segfaults. The validator dies.
Do that to leaders one at a time, and you’re not crashing a node — you’re halting the network.
Anza’s engineers learned about this from an external researcher, not a live exploit. Then they hit the real problem: publish a patch, and the diff itself reveals the bug to anyone reading it.
So they patched blind, in stages. A private message to large validators first — no details, just “a vulnerability exists, instructions follow.” Then a public hash of that message, so nobody could forge or leak it early. Only once trust was locked in did the real patch and build instructions go out, to everyone, at the same moment.
67% of the network patched within three days. Consensus was safe.
The exploit never happened. The coordination did the job the code couldn’t.
🗞️ News
The $1.3B Report That Hides a 28% Increase
CertiK’s H1 2026 Hack3D report puts total losses at $1.31 billion across 344 incidents.
Compared to H1 2025, that reads like a 47% drop.
What if last year’s number was inflated by one single hack?
It was. Bybit’s $1.45 billion wallet compromise made up 59% of all H1 2025 losses on its own.
Strip that one incident out, and H1 2025 losses were closer to $1.03 billion.
Which makes H1 2026 roughly 28% worse, not better.
The shift in attack type tells the real story.
Wallet compromise is now the single costliest category — $444.5M across just 33 incidents. Two hacks in April, Kelp DAO’s RPC compromise ($291M) and the Drift Protocol breach ($285M), account for nearly 44% of everything lost this half.
Phishing moved the opposite direction. Incident count fell over 50% year over year. Losses only fell 11%.
Attackers aren’t running fewer campaigns because they got lazier. Four targeted phishing hits made up 85% of all phishing losses in the period — fewer targets, bigger payouts.
Code vulnerabilities stayed the most common attack by count: 204 incidents, and a growing share hit contracts over a year old. Attackers are re-reading shipped code, not just auditing new deploys.
Ethereum absorbed the most incidents (153, $522.8M). Solana saw fewer hacks but outsized losses — $315M, almost entirely from Drift.
$115.3M got frozen and returned, bringing net losses to about $1.2 billion.
The headline number improved. The environment didn’t.
📚 Education
The Attack Surface Forming Before Anyone’s Watching
Agentic payments are transactions an AI agent starts and settles on its own. No human confirms anything at the moment money moves.
It’s not automation on a schedule. The agent decides what to buy, negotiates the price, and settles — in real time.
Keyrock’s 2026 report counted $73M moved across 176 million agent-initiated transactions between May 2025 and April 2026. Gartner projects agents could intermediate $15 trillion in purchases by 2028.
What happens when a bug lives in the payment rail every one of those agents shares?
It stops being one protocol’s problem. It becomes systemic.
Two camps are building the rails right now.
One is crypto-native — x402, an HTTP-native stablecoin payment protocol now under the Linux Foundation, plus the Machine Payment Protocol from Stripe and Tempo, plus Ethereum’s ERC-8183 escrow standard for agent commerce.
The other is card networks extending existing rails to machine-initiated transactions — Google handed its Agent Payments Protocol to the FIDO Alliance in April 2026, where major card networks now sit on the same working group.
Stablecoins are already winning the early volume. 98.6% of agent-settled transactions in the Keyrock data used USDC. Median transaction size: $0.01 to $0.10. Three-quarters of transactions fall below the $0.30 floor where card fees make sense at all.
Card rails were built for humans buying things a few times a day. Agents transact at machine speed, machine scale, fractions of a cent at a time — and only stablecoin settlement is cheap enough to keep up.
None of the open questions are settled. Who’s liable when an agent transacts badly? How do you audit a trail with no human in it? What limits how much an agent can spend, and on what?
The standards are still being written. The volume isn’t waiting for them.
That’s it for this week.
Reply with the Solana bug, tool, or pattern you want me to cover next — I read every one.
If a working Solana auditor in your circle would find this useful, forward it their way.
— Arsen, working Solana auditor