Vulnerability tracker
What broke, and what has to be proved again
Every break turns yesterday's attestation into history.
Attacks on trusted hardware, firmware, software supply chains, agents and robots, read with one question: which proofs made before the fix still mean something?
Intra-handshake.fail2026 · Attested TLS designs that carry attestation inside the TLS handshake
Formal analysis found relay attacks on the intra-handshake (early) attestation proposals and implementations it examined. No physical access is needed. Genuine evidence from one machine can be passed through a connection that ends somewhere else. Tracked as CVE-2026-33697 and further advisories.
What is stale: The belief that evidence verified inside the handshake proves you are talking to the attested machine.
What to prove again: Check how your attested TLS binds evidence to the connection. If it does so inside the handshake, treat the proof as open to relay until it has been reviewed. The authors note that continuous attestation is needed anyway.
Source: datatracker.ietf.org
TEE.failOctober 2025 · Intel TDX and SGX, AMD SEV-SNP, on DDR5 servers
Researchers placed a low-cost interposer on the DDR5 memory bus, recovered attestation keys and produced quotes that verify as genuine. Intel and AMD treat physical attacks as outside their threat model.
What is stale: Any decision that reads 'valid quote' as 'genuine, untampered machine' when someone can touch the hardware.
What to prove again: Add evidence of where the machine is and who operates it to the verification policy, then re-verify every node against the new policy.
Source: tee.fail
Battering RAMSeptember 2025 · Intel SGX and AMD SEV-SNP, on DDR4 servers
A cheap memory interposer that behaves during boot checks and then redirects protected memory, defeating memory encryption on cloud processors.
What is stale: Boot-time checks as proof of a trustworthy memory path for the rest of the uptime.
What to prove again: Treat physical custody as part of the attested claims. Re-attest nodes whose custody changed.
Source: batteringram.eu
WireTapOctober 2025 · Intel SGX, on DDR4 servers
A passive interposer on the memory bus let researchers extract an SGX attestation key and sign quotes of their own.
What is stale: Quotes signed under keys from machines whose physical security is unknown.
What to prove again: Limit which platforms a verifier accepts, and re-verify services that relied on self-hosted or third-party machines.
Source: wiretap.fail
BadRAMDecember 2024 · AMD SEV-SNP
Tampering with the identification chip on a memory module made the processor accept aliased memory, which broke SEV-SNP guarantees including attestation. AMD issued firmware updates.
What is stale: Every attestation report from a platform below the fixed firmware version.
What to prove again: Raise the minimum firmware version in policy. Patched nodes re-attest; unpatched nodes lose access to keys.
Source: badram.eu
CacheWarpNovember 2023 · AMD SEV-ES and SEV-SNP
A software fault attack that let a malicious host revert memory writes inside a protected virtual machine. Fixed by a microcode update.
What is stale: Reports from platforms on the old microcode level.
What to prove again: Update, raise the accepted minimum, re-attest the fleet.
Source: cachewarpattack.com
DownfallAugust 2023 · Intel processors, including SGX enclaves
A data sampling flaw that leaked data across security boundaries. Intel issued microcode and a TCB recovery for SGX.
What is stale: Enclave attestations from platforms that have not taken the update.
What to prove again: Follow the TCB recovery: new minimum version, patched platforms attest again, old ones are refused.
Source: downfall.page
xz backdoorMarch 2024 · Linux distributions shipping xz 5.6.0 or 5.6.1
A backdoor planted by a trusted maintainer reached pre-release versions of major distributions before it was found.
What is stale: 'We install the official package' as a statement of integrity.
What to prove again: Use measured software inventories to find every machine with the affected build within hours, and withdraw its reference values.
Source: nvd.nist.gov
MCP tool poisoningApril 2025 · AI agents that use Model Context Protocol tools
Hidden instructions in a tool's description can steer an agent, and a tool can change its description after the user approved it.
What is stale: An approval given to a tool whose definition has changed since.
What to prove again: Pin tool definitions, verify them at each use, and ask for approval again when they change.
Source: invariantlabs.ai
UniPwnSeptember 2025 · Unitree quadruped and humanoid robots
A flaw in the Bluetooth set-up interface allowed takeover of a robot with full privileges, and could spread from robot to robot.
What is stale: The assumption that a unit in the fleet still runs the maker's software.
What to prove again: After the fix, each unit proves its firmware state before it rejoins the fleet or receives tasks.
Source: github.com
The research behind these entries is on the reading list. How each platform is meant to work is in the platform guide. Where the field came from is in the history.
Summaries of public research, checked 12 November 2026. Read the primary source before you act. Missing an entry? [email protected]
What to do when the next one lands
- Find the units that run the affected version. Use evidence for this: a deployment plan only shows what was meant to happen.
- Change the verification policy: minimum versions, extra claims, platforms to refuse.
- Have every unit prove itself against the new policy. Units that cannot lose access to keys and data.
- Tell customers what changed, with evidence they can check.
Concepts: TCB recovery · Re-attestation · Attestation freshness · Retainer
Working on a vulnerability? A first call takes 30 minutes and commits you to nothing. Bring your product and the one question that worries you.
Next fixed date: EU Digital Identity Wallets are due in every member state on 24 December 2026, 42 days from now.