note A lot of people know me from random stuff I did on the internet, as told by random journalists who are not always great at things like nuance and fact checking. This is all mostly true, though. I am a kindergarten drop-out with authority issues, public speaking terrifies me, but this was a great excuse to see Poland and hang out with smart people, so here we are.
note I co-founded Distrust, a security consulting firm, and have had a very weird career where all of these people paid me for some reason, and several even paid me to design systems to protect hundreds of billions of dollars in financial assets. I would like to not do that anymore. And those horror stories I can never tell are why I am like this. It is really depressing and honestly a bit too easy to find security bugs in most companies, and AI can do it faster. Trying to help companies not be stupid one by one was never going to scale, but taught us a lot about what most companies struggle with in security. We still do it very selectively, and only if they really care about security. These days, most of our attention has shifted to Caution, the first general purpose verifiable compute company. Our goal is to fix the internet. We may be at it a while.
note Strip a computer apart and there are really only four places everything can go wrong, and each one is a single point of failure. The firmware the vendor ships. The supply chain that signs and builds your software. The hardware you can't really open up. And the runtime you're forced to trust all at once. I've watched every one of them get broken: LoJax and BlackLotus living quietly in firmware, xz and SolarWinds slipping through a single signing key and build farm, CPUs where we genuinely don't know what the microcode is doing, and attestation you can wave a memory-bus probe at. The uncomfortable part is that none of this gets fixed with a patch. Each one needs a structural answer at its own layer. The one I find most overlooked and easiest to exploit is the supply chain, where everything still runs like a small town where everyone leaves their doors unlocked and completely trusts each other. AI moved into town, and it is a lot harder to tell who is human these days.
note This is the trend I have had a front-row seat to my whole career. Supply chain attacks do not creep up, they jump. Every few years something bigger lands: XcodeGhost, then NotPetya hiding inside a CCleaner update, then SolarWinds, then xz. It is not getting better, it is accelerating, and AI is pouring fuel on the fire. Just to size it: Sonatype, which scans automatically and casts a very wide net, has blocked over a million malicious packages since it started counting and flagged nearly half a million new ones in a single year. The honest caveat is that the number depends a lot on who is counting, because the vendors disagree by about a hundred times. Sonatype's automated scan is the high end; Snyk, which only reports malware a human actually confirmed, counted around three and a half thousand in that same year. So the exact figure mostly tells you how wide a net was cast, not how much malware is really out there. Direction is what holds up, and every source agrees on direction. It is still a tiny fraction of the trillions of downloads out there, but when you trust a single anchor, a small percentage is all it takes. Every dot on this line is another town with its doors left unlocked.
note - Comparison table from the StageX README, ordered by objective supply-chain metrics: minimum signers, bootstrappable, reproducible. - StageX is the only row that is: 2 human signers, bootstrapped Yes, reproducible Yes, LLVM toolchain, musl + mimalloc. - The others: mostly 1 or 0 signers, single-toolchain GNU, not bootstrapped. - Refs: StageX README — https://stagex.tools
note - 181-byte **hex0 seed** → whole stack rebuilds bit-for-bit, auditable by hand in hours. - Milestones (simplified): hex0 → stage0 toolchain → mes → tinycc → gcc → llvm/musl → rustc/golang. - These 7 boxes collapse **~180 build parts** (live-bootstrap `parts.rst`) + **478 packages** — the real graph has thousands of build steps. - Built in-process — **no cross compiler to trust**; reproducible ≥2 CPU vendors, multi-signed. Nothing downloaded — derived. - Refs: hex0/Mes seeds — https://codeberg.org/oriansj/bootstrap-seeds · live-bootstrap — https://github.com/fosslinux/live-bootstrap · StageX — https://stagex.tools
note - Ladder from "anyone ships" (L0) to "defend the physical class" (L4 is the next rung, covered separately). - The **build fix (L2)** is what makes **L3 provable** and **L4 reachable**. - L3 = the Stagex enclave (Caution lands soon); L4 (open firmware/silicon) is the target, covered later. - Caution's own code is usable today at L2, but it is not yet in production. - Refs: StageX — https://stagex.tools · Caution attestation (bootproof) — https://codeberg.org/lrvick/bootproof
note - Our reference design — everything earlier is the road to it. - Firmware boots **one tiny UEFI unified kernel image (UKI) directly**, no GRUB. - The image is a **deterministic pure function**: same inputs → bit-identical output. - The supply chain + levels exist to make this **real, reproducible, verifiable** end to end. - Refs: UKI spec — https://uapi-group.org/specifications/specs/unified_kernel_image · bootproof — https://codeberg.org/lrvick/bootproof
note - Whole OS lives in one measured **UKI**: hardened offline kernel + read-only rootfs, **no shell** on the host. - **kernel → nit** (side-by-side) → **execd** (supervisor). - **execd** launches **enclaved** + three gVisors (no shell, no host network). - **enclaved** — offline service: owns TPM2 + SEV-SNP hardware; user, disk/LUKS, PCR11. - **gVisor #1** (networked execd): netstack · uutils · sshdt · rust-dhcpd · **bootproofd** — only HTTP. - **gVisor #2 (offline)** and **gVisor #3 (online)**: each runs a **User OCI Image**. - All three gVisors reach enclaved over **vsock**; remote verifier ↔ **bootproofd** over HTTP. - Status: **Preview today**, not production-ready yet. - Refs: bootproof — https://codeberg.org/lrvick/bootproof · StageX — https://stagex.tools
note - Each line: an unusual property that removes a whole class of surface, not a config toggle. - no_std/no-libc privileged core + root reduced to a **minimal Linux-capability grant** (`nit`: KEEPCAPS → drop uid/gid → no_new_privs → capset-reduce → PR_CAP_AMBIENT raise) + kernel with networking off (gVisor netstack) are the three headline ones. - Refs: notes/04-coreboot.md (kernel/nit), notes/02-stagex.md (reproducible/no download), L3 facts in the deck + research/attack-coverage.md.
note - Local = physical access to the box, defeating every software root of trust. These have no software fix today — they defeat the vendor root, not some bug you can update away. (Closing the physical class is the future goal of Level 4/open hardware, not something that ships yet.) - Glitch the AMD-SP to run unsigned SEV firmware and extract CEK/VCEK (Buhren et al. 2021); DPA on the TPM pulls the ECDH/EK key (Riscure); PSPReverse runs unsigned code on the PSP and dumps firmware — all unpatchable by firmware. TPM bus sniffing is the one that's mitigable, via parameter/bus encryption. - Refs: Buhren et al. 2021 "One Glitch to Rule Them All" · Pulse Security / SCRT / SySS TPM sniffing · Riscure TPM DPA · PSPReverse.
note - Remote = no physical access (malicious host / hypervisor / client). - Fabricked (CVE-2025-54510 / AMD-SB-3034, USENIX Security 2026): software-only, 100% deterministic — a malicious host mints SNP attestation. Supporting hostile-hypervisor CVEs: 2025-29948 (RMP bypass), 2025-48517 (SN ASID range), 2024-56161 (µcode injection), 2020-8696/8697 (SEV-ES). Fix = enforce a minimum platform TCB and reject below-TCB. - Early attestation (CVE-2026-92701/92702 + CVE-2026-33697; M. U. Sardar et al., draft-intra-handshake-fail): fails to bind evidence to the app traffic key (G3) → relay; fix = post-handshake. Partial: Cocos v0.9.0+/Privasys/CCC moved; Edgeless + Meta still vulnerable; ~24 GHSAs + EUVD-2026-16488/83192/83194. - Refs: fabricked-attack.github.io · amd.com AMD-SB-3034 · USENIX Security 2026.
note - The **exact same flow** for AMD SEV-SNP and Intel TDX — only the root-A label changes (SEV-SNP report_data·VLEK / TD quote REPORTDATA·QE-QGS). - **bootproof issues** a fresh 32-byte nonce N. - Same N signed by BOTH roots: root-A attestation report (`REPORTDATA`) **and** TPM2 quote `qualifyingData` (AK). - Report folded into slot-link **PCR22 = SHA256(report)**; quote covers PCRs 0-16 + 22. - **bootproof checks**: N in both signatures, SHA256(report)==PCR22, PCRs 0-16 == expected image digests. - Neither root sees the other's evidence -> **no replay/splice**; they agree only at the verifier. - Refs: bootproof — https://codeberg.org/lrvick/bootproof · AMD SEV-SNP attestation · Intel TDX attestation.
note - The client verifies the bundle offline against the saved file. - Two roots pinned the **same UKI image hash**; image is quorum-signed; it **mandates** disk-encryption + hardened kernel + unprivileged runtime/networking. - Note: statuses reflect the trust model — crypto rows PROVEN, image-mandated rows PROVEN. - Refs: bootproof — https://codeberg.org/lrvick/bootproof
note - These are our best-effort percentages of mitigations for each level, based on MITRE ATT&CK weighted by frequency seen in the wild, to make sure we are prioritizing the right things. - The full rationale is in the slides link at the end, and we want people to beat it up and help make it as defensible as possible before we publish the v1 Bootproof standard. - Scope: **remote, server-only (+ physical tampering, closed only at L4)**; excludes app-level, exposure, AI/ATLAS. - E2E hardware-anchored crypto → blocks **Collection / Exfiltration / much of C2**. - **No root** → PrivEsc, kernel persistence, root-level defense evasion impossible. - **No platform creds + no remote shell** → removes the credential-abuse surface (exposure opt-in; airgap supported). - L0–L3 ship today: ~5 / 10 / 30 / **95%** of in-scope surface. L4 (goal) → **100%** (incl. physical tampering, which only L4 can close); the rest is out of scope, not a gap. Weights anchored to DBIR-25 + M-Trends-25 (see research/attack-coverage.md §7). - Refs: MITRE ATT&CK — https://attack.mitre.org
note - Everything is public — links on the slide: Caution (caution.co), Distrust (distrust.co), StageX (stagex.tools), Hashbang (hashbang.sh). - Reach me: https://lance.dev · @lrvick (Mastodon) · @lrvick:matrix.org · lance@caution.co. - PGP: https://keyoxide.org/6B61ECD76088748C70590D55E90A401336C8AAA9 (fingerprint on slide) - Thank you — questions welcome.