惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

月光博客
月光博客
IT之家
IT之家
Hugging Face - Blog
Hugging Face - Blog
J
Java Code Geeks
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
博客园 - 叶小钗
MyScale Blog
MyScale Blog
G
Google Developers Blog
Microsoft Azure Blog
Microsoft Azure Blog
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
大猫的无限游戏
大猫的无限游戏
博客园 - 三生石上(FineUI控件)
Google DeepMind News
Google DeepMind News
Engineering at Meta
Engineering at Meta
The Cloudflare Blog
Martin Fowler
Martin Fowler
酷 壳 – CoolShell
酷 壳 – CoolShell
N
Netflix TechBlog - Medium
MongoDB | Blog
MongoDB | Blog
I
InfoQ
WordPress大学
WordPress大学
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
H
Help Net Security

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders
GitHub - vincentping/asys: Agentic System Interface — a t...
vincentping · 2026-06-02 · via Hacker News - Newest: "AI"

The binary system interface protocol for AI Agents — port 7816, zero shell parsing, deterministic semantics.

English | 中文

License Release Platform


Table of Contents

  • Why ASys
  • Architecture
  • Instruction Set
  • Quick Start
  • Security
  • Documentation
  • Changelog

Why ASys

SSH was designed for humans. Agents don't need a terminal.

When an AI agent runs ps aux | grep nginx over SSH, it parses free-form text that varies by OS, locale, and tool version. But when agents call ASys's SYS_PROCS instruction, they receive a fixed 44-byte binary frame: total process count, top-5 PIDs, CPU%, memory%, status flags — typed, unambiguous, the same on every node.

ASys is an experiment: what if you designed a system interface specifically for AI agents, from first principles? Binary frames instead of text. A long-lived encrypted connection instead of per-command sessions. Instruction-level capability grants instead of broad SSH access. Built-in audit trail instead of shell history.

It's not a replacement for SSH, Ansible, or Kubernetes operators — those tools are well-suited for their intended users (humans and orchestration pipelines). ASys is an additional option for the case where the operator is an AI agent and you want an interface designed for that from the ground up.

To understand the full design rationale and where ASys fits in the agent infrastructure landscape, start with the whitepaper.


Architecture

  AI Agent (LLM)
       │
       │  Tool calls
       ▼
  Python SDK  (~/.asys/id_curve25519)
       │
       │  Noise_IK_25519_ChaChaPoly_BLAKE2b
       │  TCP port 7816
       ▼
  asyd  (C daemon, zero dependencies)
       │
       │  POSIX syscalls
       ▼
  Linux

Instruction Set

Core ISA (0x00–0x0F) — read-only, zero side effects

INS Name Response Description
0x00 SYS_CAPS 36B Static capabilities: CPU, RAM, disk, ISA bitmap
0x01 SYS_HELLO 18B Node UID + nanosecond timestamp
0x02 SYS_STATUS 23B Load avg, CPU%, available RAM, disk, network, RPI
0x03 SYS_PROCS 44B Total procs + top-5 by CPU (PID, CPU%, MEM%)

Protocol Control (0x10–0x1F)

INS Name Response Description
0x11 TASK_QUERY 3B Poll async task status by Task_Handle

Standard ISA — Process Control (0x20–0x2F) — signed, require elevated capabilities

INS Name Response Description
0x20 PROC_THROTTLE SW only SIGSTOP / SIGCONT a process by PID
0x21 NET_ISOLATE Isolate process network access (planned)
0x22 SVC_RESTART 6B Restart a named systemd service (async)

Measured RTT (Noise IK encrypted, RHEL on VirtualBox, Windows Python client):

Instruction RTT Notes
SYS_HELLO < 1ms
SYS_CAPS < 1ms No cache; static data read once at startup
SYS_STATUS < 1ms / ~51ms Cache hit / cold sample (50ms CPU dual-sample)
SYS_PROCS ~6ms / ~200ms Warm calls; first call blocks 200ms cold sample
PROC_THROTTLE ~200µs dispatch
SVC_RESTART ~200µs dispatch Async; poll result with TASK_QUERY

Quick Start

Prerequisites

  • Linux (RHEL/Fedora/Ubuntu/Debian), x86_64
  • gcc, make
  • Python 3.8+ with pip install noiseprotocol cryptography (client only)

Build and run

git clone https://github.com/vincentping/asys
cd asys
make
sudo bin/asyd

asyd listens on TCP 7816. On first start it generates a key pair at /etc/asyd/id_curve25519.

Options:

Option Default Description
--port <n> 7816 TCP listen port
--listen <addr> 0.0.0.0 Bind address
--debug off Run in foreground with verbose logging to stderr
--version Print version and exit
--help Print usage and exit
# Example: custom port, local-only
sudo bin/asyd --port 8816 --listen 127.0.0.1

# Example: debug mode (foreground, verbose)
sudo bin/asyd --debug

Register an agent

# On the client machine — generate agent identity
python3 tools/client/asys_keygen.py

# On the server — add the agent's public key to the whitelist
echo "<pubkey_hex>" | sudo tee -a /etc/asyd/authorized_agents

The agent public key is printed by asys_keygen.py. It also generates the client key pair at ~/.asys/id_curve25519 used during the Noise IK handshake. Connections from agents not in /etc/asyd/authorized_agents are rejected with SW=0x6982.

On first connection, the client will prompt you to confirm the server's public key fingerprint and save it to ~/.asys/known_hosts — see First connection below.

Reload the whitelist without restarting (existing connections are not affected):

sudo kill -HUP $(pidof asyd)

First connection

On first connect, the client displays the server's public key fingerprint:

ASys server fingerprint (SHA256): a3:f1:2c:...
Connect to localhost:7816? [yes/no]: yes
Fingerprint saved to ~/.asys/known_hosts

Subsequent connections verify the fingerprint automatically. A mismatch aborts the connection — same model as SSH known_hosts.

Run the demos

All client scripts run on any machine (including Windows) and connect to a remote asyd over the network. No SSH. No shell. Signed binary instructions over an encrypted channel.

1. Core ISA — verify the full connection stack

Connects to asyd, completes the Noise IK handshake, then exercises all four Core ISA instructions (SYS_CAPS, SYS_HELLO, SYS_STATUS, SYS_PROCS) and runs a cache timing test.

python3 examples/client_core_isa.py <server-ip>
ASys Test Client  —  localhost:7816
====================================================
Connected to localhost:7816
Handshake OK

── SYS_CAPS (0x00) ─────────────────────────────────
  Core bitmap:      0x0002000F  active=['0x00', '0x01', '0x02', '0x03', '0x11']
  Ext  bitmap:      0x00000005
  Protocol:         v1.0
  Kernel hash:      0x06060057
  CPUs:             8  arch=x86_64
  RAM:              15660 MB   swap=4096 MB
  Root partition:   1031018 MB   fs=ext4
  RPI type:         NATIVE_KERNEL (PSI)
  SW:          0x9000  OK

── SYS_HELLO (0x01) ────────────────────────────────
  Magic:            'ASYS'
  Node UID:         0xFCAB032F
  Server timestamp: 1779912196.304 s  (1779912196304227300 ns)
  RTT:              0.28 ms
  SW:          0x9000  OK

── SYS_STATUS (0x02) ───────────────────────────────
  Load avg:         1m=0.0  5m=0.0
  CPU usage:        0%
  Mem available:    14540 MB
  Root disk:        0% used   IO wait=0 ms
  Network:          in=0 Mbps   out=0 Mbps
  RPI:              0/100
  RTT:              51.05 ms
  SW:          0x9000  OK

── SYS_PROCS (0x03) ────────────────────────────────
  Total processes:  48
  Top 5 by CPU (lifetime share):
    [0] PID=5964    CPU=   0.07%  MEM=  0%  flags=Privileged
    [1] PID=433     CPU=   0.05%  MEM=  0%  flags=Zombie
    [2] PID=1       CPU=   0.02%  MEM=  0%  flags=Privileged
    [3] PID=6       CPU=   0.00%  MEM=  0%  flags=Privileged
    [4] PID=80      CPU=   0.00%  MEM=  0%  flags=Privileged
  RTT:              2.87 ms
  SW:          0x9000  OK

── Cache Timing Test ───────────────────────────────
   6 rounds, 200ms interval
   #       SYS_STATUS     SYS_PROCS
   ────  ────────────  ────────────
   0           0.61ms        2.52ms
   1           0.85ms        2.92ms
   2           0.95ms        2.47ms
   3          52.25ms        2.68ms
   4           0.95ms        2.68ms
   5           1.21ms        2.70ms

====================================================
Done.

2. SVC_RESTART — async instruction pattern

Sends a SVC_RESTART instruction, receives a Task_Handle, then polls TASK_QUERY until the service restart completes.

python3 examples/client_svc_restart.py <server-ip> 7816 sshd
ASys Phase 3 E2E Test  —  localhost:7816
Service: sshd
====================================================
Handshake OK

► SVC_RESTART("sshd")
  SW:           0x9000  OK
  Task_Handle:  0x001D7D05
  RTT:          1.0 ms

► TASK_QUERY polling (max 30s) ...
  [ 1s]  Status = Success ✓

====================================================
PASS  SVC_RESTART("sshd") completed with Status=Success

3. Multi-agent — concurrent session isolation

Spawns four independent agents concurrently to validate per-session locking: concurrent reads don't interleave, cross-session TASK_QUERY leaks no handle information, and an abrupt disconnect does not affect other sessions.

python3 examples/client_multi_agent.py <server-ip>
=== ASys Agent Simulator — localhost:7816 ===
Validates per-session lock correctness (Phase 4 P1)


[Scenario 1: Concurrent Core ISA — 4 agents × 5 SYS_STATUS]
  PASS  Agent-A2: 5× SYS_STATUS all SW=0x9000
  PASS  Agent-A3: 5× SYS_STATUS all SW=0x9000
  PASS  Agent-A1: 5× SYS_STATUS all SW=0x9000
  PASS  Agent-A4: 5× SYS_STATUS all SW=0x9000

[Scenario 2: Cross-session TASK_QUERY isolation]
  PASS  Agent-B1: SVC_RESTART crond → SW=0x9000
  PASS  Agent-B1: obtained valid Task_Handle
  PASS  Agent-B2 querying B1's handle → Status=0xFF (no info leak)
  PASS  Agent-B1 querying own handle → Status != 0xFF

[Scenario 3: Concurrent SVC_RESTART — 3 agents, 3 different services]
  PASS  Agent-C1: SVC_RESTART crond   → SW=0x9000 + handle
  PASS  Agent-C2: SVC_RESTART rsyslog → SW=0x9000 + handle
  PASS  Agent-C3: SVC_RESTART sshd    → SW=0x9000 + handle
  PASS  All 3 Task_Handles are distinct
  FAIL  Agent-C1: crond   → Status=Success ✓  (Status=Failed)
  FAIL  Agent-C2: rsyslog → Status=Success ✓  (Status=Failed)
  PASS  Agent-C3: sshd    → Status=Success ✓

[Scenario 4: Disconnect resilience — Agent-D abruptly disconnects]
  PASS  Agent-D-stable: SYS_HELLO before disrupt → 0x9000
  PASS  Agent-D-bad: abrupt disconnect (no handshake)
  PASS  Agent-D-stable: SYS_STATUS after disrupt → 0x9000
  PASS  Agent-D-new: fresh connect after disrupt → SYS_HELLO 0x9000

====================================================
Summary: 17 passed, 2 failed
SOME TESTS FAILED — check output above

Note on Scenario 3 failures: Status=Failed for crond/rsyslog means systemctl restart returned non-zero — those services are not installed or not running on the test node. The ASys protocol path (dispatch → fork → SIGCHLD → handle update) is exercised correctly regardless; Status=Failed is the expected response when the underlying system operation fails.

4. PROC_THROTTLE — observe and control a live process

The demo uses two machines to show what ASys is actually for: a remote operator observing and controlling a Linux node over the network.

On the server (RHEL/Linux) — start a CPU hog:

python3 examples/server_cpu_hog.py
CPU hog started  PID=2644
Press Ctrl+C to stop.

On the client (any machine, e.g. Windows) — connect and throttle:

python3 examples/client_proc_throttle.py <server-ip>
ASys Throttle Client  —  <server-ip>:7816
====================================================
Handshake OK

  CPU=100%  load1m=3.2  RPI=87/100

  #    PID       CPU%   MEM%
  ---  --------  ------  -----
  1    2644      99.87%     0%
  2    1281       9.09%     0%
  3       1       0.00%     1%
  4       9       0.00%     0%
  5      17       0.00%     0%

Select # or PID to throttle (blank to cancel): 1
  Selected #1 → PID 2644

  PROC_THROTTLE STOP  PID=2644 ...
  SW=0x9000  OK
  Waiting 2 s for CPU to settle...
  CPU=0%  (was 100%,  Δ=-100%)

  PID 2644 paused.  Press Ctrl+C to restore or exit.
^C
  Restore PID 2644? (yes/no) [no]: yes
  PROC_THROTTLE CONT  PID=2644  SW=0x9000  OK

The client runs on Windows. The process being throttled is on RHEL. No SSH. No shell. One signed binary instruction over an encrypted channel.


Security

Transport: Noise IK (Curve25519 + ChaCha20-Poly1305 + BLAKE2b) — 1-RTT mutual authentication. No passwords. No certificates. No CA. Session content has forward secrecy (session keys are derived from per-handshake ephemeral keys; compromising the static private key does not expose past session content). Known boundary: the agent's static public key is encrypted under the server's static public key during the handshake — a static private key compromise would theoretically allow recovery of agent identities from recorded handshake traffic.

Server identity: on every new connection, asyd sends a 38-byte Pre-Handshake Frame ([Magic][Version][ServerPubKey]) before the Noise handshake. The client verifies the server's public key fingerprint against ~/.asys/known_hosts (SSH-style: confirm on first connect, reject on mismatch thereafter). The public key is public by definition; security relies on fingerprint confirmation, not secrecy.

Replay protection: signed instructions carry an Epoch_ID (HMAC-BLAKE2b(recv_key, "asys-epoch-v1")[:4]), derived post-handshake and never transmitted. Each session produces a unique Epoch_ID; cross-session replay is cryptographically impossible without breaking the session key. Within a session, a monotonic sequence number (Seq) prevents replay with 0x6985.

Privilege containment: asyd runs as Caged Root under systemd: CapabilityBoundingSet limits privileges to exactly what each instruction needs (CAP_KILL, CAP_SYS_RESOURCE, CAP_NET_ADMIN, CAP_SYS_PTRACE, CAP_DAC_READ_SEARCH). NoNewPrivileges=true prevents escalation even if the process is compromised.

Instruction security levels (CLA byte, bits 3–2):

Level Applies to Mechanism
Plain Core ISA Noise channel encryption only
Signed Standard ISA HMAC-BLAKE2b Auth Tag per frame

Documentation

Document Description
docs/en/asys-whitepaper.md Background, design rationale, available options, and where ASys fits
docs/en/asys-spec.md Protocol specification: ISA, security model, APDU frame format
docs/en/asys-design-notes.md Architecture decision records (why not JSON / mTLS / shell)
docs/en/asys-conformance.md Conformance testing guide
sdk/definitions/asys.isa Machine-readable ISA definition (TOML)

Changelog

Version numbers: asyd software versions (v0.3.x) and the ASys protocol version (v1.0, 0x0100) are managed independently. The protocol version increments only when the wire format or instruction semantics change.

v0.3.1 — 2026-05-28

  • Client-Speak-First (ADR-22): after TCP connect, asyd waits for the client to send a 4-byte Magic (0x41535953) before sending anything; 1-second hardcoded timeout; mismatch or timeout → silent close, no response. Reduces exposure to generic port scanners.
  • --help and --version flags added to asyd
  • Conformance: test_client_magic.c (TC-MAG-001/002/003) added to make check

v0.3.0 — 2026-05-27 (initial open-source release)

  • asyd C daemon, zero external dependencies, static memory pool, zero malloc on the request path
  • Noise IK (Curve25519 + ChaCha20-Poly1305 + BLAKE2b) transport, 1-RTT mutual authentication
  • Agent public key whitelist (/etc/asyd/authorized_agents), SSH-style fingerprint verification
  • Pre-Handshake Frame: server broadcasts public key before handshake, ~/.asys/known_hosts management
  • HMAC-BLAKE2b Auth Tag verification + Epoch_ID cross-session replay protection
  • Core ISA: SYS_CAPS, SYS_HELLO, SYS_STATUS, SYS_PROCS
  • Standard ISA: PROC_THROTTLE, SVC_RESTART (async, with TASK_QUERY)
  • Per-session concurrency, 60s idle timeout, journald logging, systemd Caged Root deployment

License

Apache 2.0 — Copyright (c) 2026 Vincent Ping (vincentping@gmail.com)

Monocypher (cryptography library): CC-0 / ISC dual license.