Confidential Computing Attack logo

Confidential Computing Attack

Community
brucesongs
confidential-computing-attack

Attacks against Trusted Execution Environments (TEEs) and confidential computing platforms — Intel SGX (Foreshadow/SGAxe/LVI/ÆPIC Leak), Intel TDX, AMD SEV/SEV-ES/SEV-SNP (CrossLine/BadRAM), Azure CCF, Marblerun, and Gramine/Occlum libos enclaves. Covers side-channel leakage, attestation forgery, ABI misuse, host-to-enclave breakout, enclave-to-host escape, and recovery of sealed secrets. Distinct from hardware-security (broad hardware attacks) and firmware-reverse (UEFI/BIOS).

Overview

Publisherbrucesongs
Repositorykali-claw
Skill nameconfidential-computing-attack
Stars
70
Forks
18
Bundled files
9
LicenseMIT
Links
  • Markdown instructions

    A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.

  • Works with any LLM

    AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.

  • 9 bundled files

    Scripts, templates, and references the model can read while it works. Files are read-only and never executed.

  • Open source

    Published by brucesongs on GitHub. Read the source before you install it.

Installation

Install the Confidential Computing Attack AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

1

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

  1. Open the app and go to Plugins → Skills.
  2. Choose "Install from GitHub".
  3. Paste the skill folder URL below and confirm.
  4. Enable the skill in any chat where you want it available.
Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.
2

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills
git clone --depth 1 https://github.com/brucesongs/kali-claw.git /tmp/kali-claw
mkdir -p .claude/skills
cp -r /tmp/kali-claw/skills/confidential-computing-attack .claude/skills/confidential-computing-attack
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Confidential Computing Attack in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Confidential Computing Attack on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Confidential Computing Attack is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

  • Loaded only when it is needed

    The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.

  • Switch models mid-chat

    Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

Confidential Computing Attack

Supplementary Files:

  • payloads.md — Attack payloads organized by TEE: Intel SGX (Foreshadow/SGAxe/LVI/ÆPIC Leak), Intel TDX, AMD SEV-SNP (CrossLine/BadRAM), Azure CCF, Marblerun, Gramine/Occlum
  • test-cases.md — 18 structured test cases covering attestation forgery, side-channel leakage, ABI misuse, host-to-enclave breakout, and sealed-secret recovery
  • guides/confidential-computing-attack-playbook.md — End-to-end playbook with engagement scoping, lab setup (QEMU with SGX/SEV emulation), attestation workflow attacks, enclave breakout methodology, and blue-team detection
  • guides/real-world-incident-case-studies.md — Eight case studies including SGX Foreshadow (2018), SGAxe (2020), LVI (2020), ÆPIC Leak (2022), AMD SEV CrossLine (2025), BadRAM (2024), and Gramine ABI abuse

Summary

This skill targets confidential computing — the technology that protects data in use via hardware-isolated enclaves. While at-rest and in-transit encryption are well-understood, TEEs introduce a new attack surface: enclave attestation forgery, microarchitectural side channels, ABI misuse between enclave and untrusted host, and host-to-enclave breakout.

Tools: sgx-step, sgvisor, sev-tool, snp-lab, marblerun-cli, ccf-client, occlum-tools, gramine-direct, rust-sgx, edgelessrt, intel-sgx-sdk, sgxs-tools, libsgx, sgx_emulate

Domain: confidential-computing

MITRE ATT&CK: T1046 Network Service Discovery · T1602 Data from Configuration Repository · T1068 Exploitation for Privilege Escalation · T1556 Modify Authentication Process

Description

Attacks against Trusted Execution Environments (TEEs) and confidential computing platforms — Intel SGX (Foreshadow/SGAxe/LVI/ÆPIC Leak), Intel TDX, AMD SEV/SEV-ES/SEV-SNP (CrossLine/BadRAM), Azure CCF, Marblerun, and Gramine/Occlum libos enclaves.

Confidential computing assumes a strong threat model: the OS, hypervisor, BIOS, and even physical DRAM are untrusted. Only the TEE itself (the CPU enclave) is trusted. Attacks against this model fall into four broad categories:

  1. Microarchitectural side channels — exploiting CPU cache, branch predictor, TLB, or memory bus behavior to leak enclave secrets
  2. Attestation forgery — bypassing or faking attestation reports to make untrusted code appear trusted
  3. ABI misuse — abusing the carefully-specified boundary between enclave and untrusted host to corrupt enclave state
  4. Host-to-enclave breakout — escaping from an enclave back to the host, or vice versa

This skill is distinct from:

  • hardware-security — which targets physical hardware attacks (JTAG, side-channel on chips, fault injection)
  • firmware-reverse — which targets UEFI/BIOS, not runtime enclaves
  • container-security — which targets container runtimes, not hardware isolation
  • cloud-security — which targets cloud control planes broadly

Where container-security asks "can I escape this container?", confidential-computing-attack asks "given that this enclave is supposed to be impenetrable even to a malicious cloud provider, what corner cases let me violate that promise?"

Use Cases

  • SGX attestation forgery — craft an IAS/DCAP attestation report that appears valid but represents untrusted code (e.g., for OCSP relay attacks)
  • SGX sealed-secret recovery — extract the Seal Key from a compromised platform to decrypt sealed secrets offline
  • Foreshadow (L1TF) attack — read enclave memory from a sibling thread via L1 Terminal Fault
  • SGAxe attack — extract the platform's EPID group key, allowing attestation forgery for any enclave on that platform
  • LVI attack — inject faulting load operations to hijack enclave execution
  • ÆPIC Leak — exploit APIC MMIO read to leak SGX enclave data
  • AMD SEV-SNP attestation forgery — tamper with VLEK/VCEK signed attestation reports
  • AMD CrossLine attack — break VM-to-VM isolation within SEV-SNP
  • BadRAM attack — exploit DDR5 SPD SPD-Bypass on consumer DIMMs to break SEV-SNP guest memory integrity
  • TDX attestation abuse — manipulate TDREPORT to make untrusted VM appear as trusted
  • Azure CCF governance attack — abuse the consortium governance model to add a malicious node
  • Marblerun manifest tamper — exploit weak EManifest validation to alter marble policy at runtime
  • Gramine/Occlum libos escape — exploit syscall-emulation bugs to escape the libos back to host

Differentiation from Adjacent Skills

SkillTargetBoundaryTypical entry
hardware-securityPhysical chips, IoT devicesPhysical accessJTAG, glitching
firmware-reverseUEFI/BIOS, Option ROMBoot-timeDMA attacks, SMM
container-securityOCI, K8s, containerdKernel namespaceContainer escape
confidential-computing-attack (this)TEEs (SGX, TDX, SEV-SNP, CCF)CPU enclave boundarySide channel, attestation abuse
cloud-securityCSP control planeCloud IAMIAM enumeration
kernel-exploitationOS kernelUser/kernel boundarySyscall misuse

Core Tools

Intel SGX

  • Intel SGX SDK — official SDK; includes sgx_sign, sgx_edger8r
  • sgx-step — single-step enclave execution, page-fault side channels (Ard et al., TU Wien)
  • sgvisor — enclave hypervisor framework for research
  • sgxs-tools — Rust utilities for SGX enclave file manipulation
  • Foreshadow PoC — academic PoC for L1TF (Wikipedia: Foreshadow)
  • SGAxe PoC — academic PoC for EPID key extraction
  • LVI PoC — academic PoC for Load Value Injection
  • ÆPIC Leak PoC — CVE-2022-21233 PoC

Intel TDX

  • tdx-tools — Intel TDX development kit
  • tdquote — TDX attestation quote generation
  • QEMU with TDX support — confidential VM testing

AMD SEV / SEV-ES / SEV-SNP

  • amdSEVTool — SEV/SEV-SNP attestation utilities
  • sev-snp-lab — local CVM testing harness
  • SEV-SNP guest tools — guest-side attestation helper
  • CrossLine PoC — research tooling for SEV-SNP VM isolation attacks

Azure CCF

  • ccf-cli — CCF consortium governance CLI
  • ccf-client — RPC client for CCF nodes
  • ccf-recovery — disaster recovery tooling

Marblerun

  • marblerun-cli — manifest management
  • edgelessrt — runtime SDK

LibOS Enclaves

  • gramine-direct — Gramine libos CLI
  • occlum-tools — Occlum libos CLI
  • sgx-lkl — SGX-LKL libos

Methodology

Phase 1 — Reconnaissance

  • Identify SGX/TDX/SEV-SNP usage in target (binary inspection for SGX instructions like encls, enclu, enclv)
  • Identify attestation provider (Intel IAS, Intel DCAP, AMD VCEK, Azure CCF, custom attestation service)
  • Map enclave binary: identify ECALLs, OCALLs, edge routines
  • Identify attestation report format (EPID, DCAP, ECDSA)
  • Identify sealed-secret storage locations

Phase 2 — Attestation Attacks

  • Replay old attestation reports
  • Forge attestation reports via IAS/DCAP compromise
  • Exploit weak quote verification (date, nonce, key)
  • Abuse advisory (allow-list) policies on outdated enclave versions
  • Apply CCF governance attack to add malicious node

Phase 3 — Microarchitectural Side Channels

  • Identify secret-dependent branches in enclave
  • Deploy cache timing attacks (Prime+Probe, Flush+Reload)
  • Deploy branch prediction attacks (Spectre variant in enclave)
  • Deploy SGX-Step for single-step page-fault side channels
  • Apply Foreshadow (L1TM) for enclave memory reads
  • Apply SGAxe for EPID key extraction

Phase 4 — ABI Misuse

  • Review ECALL/OCALL signatures for untrusted-pointer dereferences
  • Identify buffer overflows in edge routines
  • Identify integer overflows in array length checks
  • Exploit missing check_*
  • Race-condition in asynchronous edge calls (AEX race)

Phase 5 — Enclave Breakout / Escape

  • Identify libos syscall-emulation bugs (Gramine, Occlum)
  • Exploit memory layout assumptions (TLS region, stack)
  • Forge shared memory mappings between enclave and host
  • Abuse insecure key derivation for sealed secrets

Phase 6 — Sealed Secret Recovery

  • Extract Seal Key derivation parameters from platform
  • Compute Seal Key offline using leaked CPU_SIG or fused key
  • Decrypt sealed secrets from disk
  • Pivot to enclave's persisted state

Phase 7 — TEE Host Compromise

  • Compromise host kernel to feed malicious SGX instructions
  • Use AEX notify to time enclave execution
  • Deploy ÆPIC Leak (CVE-2022-21233) for APIC-based read
  • Deploy LVI (CVE-2020-0551) for faulting-load injection

Phase 8 — TDX / SEV-SNP Specific

  • TDX: abuse TDREPORT signing
  • SEV-SNP: exploit VCEK key reuse
  • SEV-SNP: apply CrossLine for VM-to-VM isolation break
  • SEV-SNP: apply BadRAM for DDR5 SPD-bypass memory integrity break

Defense Perspective

Core Principles

  1. Defense in depth — assume any single TEE may be compromised; layer attestation, code signing, and runtime checks
  2. Quote verification is critical — never accept attestation reports without verifying signature, freshness, and nonce
  3. Advisory policy is policy — Intel's "advisory allowlist" lets unpatched enclaves still attest; decide explicitly
  4. Sealed secrets require key rotation — rotate Seal Key after platform firmware updates
  5. Side-channel resistance is hard — assume data-dependent branches leak; use constant-time crypto

Hardening Checklist

  • Require attestation quote verification (signature, nonce, freshness) at every RPC
  • Apply Intel advisory policy explicitly (e.g., reject reports with vulnerable advisory IDs)
  • Use ECDSA attestation (DCAP) over EPID where possible
  • Use SEV-SNP over SEV/SEV-ES for new deployments
  • Apply page-fault side-channel defenses (occlum/gramine mitigations)
  • Use constant-time crypto primitives (libsodium, OpenSSL CT)
  • Rotate Seal Keys after firmware updates
  • Audit attestation reports in tamper-evident log (CCF)

Detection

  • Attestation report anomaly detection (new advisory IDs, new MRENCLAVE)
  • Side-channel detection via performance counter anomalies
  • Sealed-secret decryption anomaly rate
  • TEE error injection attempts (enclave CPUID mismatch)

Practical Steps

SGX Recon

  1. nm enclave.signed.so | grep -E "sgx_|ecall" to find entry points
  2. objdump -d enclave.signed.so for SGX instruction patterns (enclu, encls)
  3. sgx_sign dump -key ... -enclave ... -dumpfile ... for metadata

SGX Attestation Replay

  1. Capture a valid attestation report (older or from compromised platform)
  2. Replay to relying party without fresh nonce
  3. If relying party doesn't verify nonce → accept stale report

SGX Foreshadow (L1TM) Recovery

  1. Deploy sibling thread that triggers enclave page fault
  2. Use sgx-step to single-step enclave execution
  3. Apply L1 Terminal Fault to leak enclave cache lines
  4. Reconstruct secret from cache line samples

AMD SEV-SNP Attestation Audit

  1. sev-tool --get_vcek to fetch the platform's VCEK certificate
  2. Verify attestation report signature against VCEK
  3. Check policy fields (abi_major, abi_minor, smt)
  4. Compare MEASUREMENT field against expected guest measurement

Azure CCF Governance Attack

  1. Identify network consortium (config.json: members, certificates)
  2. Compromise a member's private key (or social-engineer a quorum member)
  3. Submit set_constitution proposal to add malicious operator
  4. If accepted → malicious node joins network

Gramine LibOS Recon

  1. gramine-sgx-sign --manifest app.manifest to inspect manifest
  2. gramine-sgx-token to generate runtime measurement
  3. Identify syscall emulation bugs via fuzzing (e.g., gramine-sgx-direct --trace-syscalls)

Detection Methods

TEE Audit Logging

  • SGX attestation logs: Failed attestation attempts; quote verification failures.
  • AMD SEV-SNP logs: VMPL switch anomalies; TCB mismatch events.
  • Intel TDX logs: TD exit reasons; malicious TD guest indicators.

Side-Channel Detection

  • Cache timing anomalies: Process exhibiting consistent timing patterns (signature of cache attack).
  • Page fault patterns: Single process causing high page fault rate on shared memory.
  • Power analysis signatures: Hardware-level monitoring detects DPA/CPA attack patterns.

SIEM Detection Rules

  • Splunk SPL: index=tee sourcetype=sgx_attest | where status="failed" | stats count by enclave_id
  • Custom HSM monitoring: Detect anomalous command sequences to HSM (key extraction attempts).

Defense Evasion Techniques

Attestation Bypass

  • Attestation spoofing: Replay legitimate attestation quote (some implementations don't bind nonce to enclave state).
  • Outdated TCB: Target enclaves with old TCB; vulnerabilities known but not patched.
  • Side-channel tolerance: Use tolerated side-channels (Foreshadow, Spectre); not flagged by enclave.

TEE Escape Stealth

  • CVE selection: Use newer CVEs (CVE-2022-40982, CVE-2023-23583) before detection rules updated.
  • Cross-enclave attacks: Use compromised enclave to attack others; appears as legitimate communication.
  • Memory disclosure via side-channel: No direct vulnerability exploited; harder to attribute.

Cloud Confidential VM Attacks

  • Snapshot abuse: Snapshot CVM disk, restore outside CVM, mount to read encrypted data.
  • VBS (Virtualization-Based Security) escape: Target Windows VBS flaws; bypass Credential Guard.
  • SEV-SNP VMPL confusion: Abuse VMPL levels to escalate privileges within SEV-SNP guest.

Cross-References

  • skills/hardware-security/SKILL.md — physical hardware attacks (JTAG, fault injection)
  • skills/firmware-reverse/SKILL.md — UEFI/BIOS / SMM attacks
  • skills/binary-reverse/SKILL.md — enclave binary static analysis
  • skills/crypto-attacks/SKILL.md — attestation report cryptography analysis
  • skills/cloud-security/SKILL.md — confidential VMs in cloud (Azure DCsv3, AWS Nitro Enclaves)
  • skills/container-security/SKILL.md — confidential containers (Kata + SGX)

References

  • Intel — Intel SGX Developer Reference (2024)
  • AMD — SEV-SNP Firmware ABI Specification (2024)
  • Microsoft — Azure Confidential Cloud (CCF) Documentation (2024)
  • Foreshadow — Masters of SGX (USENIX Security 2018)
  • SGAxe — How to Steal the SGX Attestation Key (CCS 2020)
  • LVI — Load Value Injection (IEEE S&P 2020)
  • ÆPIC Leak — CVE-2022-21233 (BlackHat USA 2022)
  • CrossLine — Breaking SEV-SNP VM Isolation (USENIX Security 2025)
  • BadRAM — DDR5 SPD Bypass (USENIX Security 2024)
  • Gramine — Documentation and Threat Model (2024)
  • Occlum — Threat Model Paper (IEEE S&P 2021)

Bundled files

The model reads these on demand while the skill is loaded. They are exposed as readable files and are never executed.

Frequently asked questions

What does the Confidential Computing Attack AI skill do?

Attacks against Trusted Execution Environments (TEEs) and confidential computing platforms — Intel SGX (Foreshadow/SGAxe/LVI/ÆPIC Leak), Intel TDX, AMD SEV/SEV-ES/SEV-SNP (CrossLine/BadRAM), Azure CCF, Marblerun, and Gramine/Occlum libos enclaves. Covers side-channel leakage, attestation forgery, ABI misuse, host-to-enclave breakout, enclave-to-host escape, and recovery of sealed secrets. Distinct from hardware-security (broad hardware attacks) and firmware-reverse (UEFI/BIOS).

Why use Confidential Computing Attack on TypingMind?

Because you install it once and use it with any model. Confidential Computing Attack is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Confidential Computing Attack in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/brucesongs/kali-claw/tree/main/skills/confidential-computing-attack. TypingMind reads its SKILL.md and bundles its files and installs it as a skill you can enable per chat.

Which AI models can use Confidential Computing Attack?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Confidential Computing Attack?

As many as you like. As long as a model supports skills, you can use Confidential Computing Attack with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Confidential Computing Attack AI skill free?

Yes. It is published on GitHub by brucesongs under the MIT license. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Set up your own AI workspace now

Get notified about new features and future giveaways by subscribing to our newsletter 👇