Previous editions
AgentLens

Agent Security Weekly

Weekly threat briefing

Stolen keys put agents at risk

TL;DR

  • Cursor says a compromised key at AI provider Fireworks led to malicious Claude Code extensions, and the breach is still the clearest supply-chain case this month Dark Reading.
  • Google describes attackers running autonomous multi-agent tools, but the Java and "first fully autonomous attack" headlines are weak leads The Hacker News dev.ua Cyber Magazine.
  • Supergateway and vLLM both have new CVEs where an unauthenticated or hostile client can reach the agent or model backend CVE-2026-132121 CVE-2026-130686 CVE-2026-130685.

The week in agent security

This week the risk sits at the edges of the agent, not inside the model. A stolen vendor key reached Cursor data and credentials Dark Reading. A researcher group took over a popular model account on Hugging Face because the platform does not verify ownership BleepingComputer. In both cases, trust in a third party was the weak point.

The same pattern shows up in serving software. Supergateway lets a web page talk to a local MCP process CVE-2026-132121. Two vLLM bugs let remote clients exhaust resources or crash the engine CVE-2026-130686 CVE-2026-130685. These are classic web bugs in new places. In our view, many AI tools were built for a trusted laptop, then exposed to networks and browsers.

On the attacker side, Google says threat actors now use autonomous frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to build and run exploits from patches The Hacker News. Reports on ransomware and a phishing kit add that AI mostly gives scale and better lures Dark Reading Dark Reading. The defence answer is the same as always: limit what a stolen key or a steered agent can reach. Simon Willison's work on external sandboxes follows that idea Simon Willison.

Top stories

Supply chain

Fireworks key leads to malicious Claude Code extensions

Cursor traced malicious Claude Code extensions to a compromised key issued to Fireworks. Threat actors used it to reach Cursor data and credentials Dark Reading. The excerpt does not say what the extensions did or how many people installed them. Anyone who installed them should treat local credentials as exposed.

Threat level: High. Real compromise with a confirmed path, but payload details are missing.

Tool security

Supergateway cross-site WebSocket hijacking

CVE-2026-132121 affects Supergateway 2.4.0 through 4.2.2 in stdio-to-WebSocket mode. It does not check the Origin header, so a malicious page can open the /message WebSocket from the victim's browser CVE-2026-132121. In our view, that could let a web page drive a local MCP server and its tools.

Threat level: High. Public details on a real MCP bridge; upgrade or restrict access.

Model security

Hugging Face account takeover of Kokoro

Researchers took control of the account for Kokoro, a text-to-speech model used in projects from Google, NVIDIA, AMD, Intel and Amazon Web Services. The cause was missing ownership verification BleepingComputer. Anyone who pulls models by name could be served changed files.

Threat level: Medium. Done by researchers, but the weakness is real and the reach is wide.

vLLM resource exhaustion and engine crash

CVE-2026-130686 lets remote clients exhaust frontend resources with many image URLs in one request CVE-2026-130686. CVE-2026-130685 lets unauthenticated attackers crash EngineCore with out-of-range token ids when --skip-tokenizer-init is used CVE-2026-130685. Both are fixed in 0.33.0 per the excerpts.

Threat level: Medium. Denial of service only, but it affects shared inference hosts.

Expert take: threats and trends

Compared with last week, the Fireworks breach continued, and its main facts are unchanged Dark Reading. Supergateway continued as a story, now as a known CVE rather than a new find CVE-2026-132121. Google's autonomous agent report continued, and no new frameworks were named The Hacker News. Uber's victim count did not appear this week.

New this week: model-hosting trust (Hugging Face) and inference server bugs (vLLM). Attackers are going after the supply chain around agents, because it is easier than attacking the model.

Over-hyped: the "AI found a critical Java flaw threatening 90% of top companies" headlines. These are search leads with no detail, so we do not rate them dev.ua Anadolu Ajansı. Same for "first fully autonomous AI attacks" Cyber Magazine and the vendor pentest pieces, which are marketing-adjacent The Hacker News The Hacker News. The prompt injection explainer is background only qz.com. We skip the currency items as noise link link.

Watch next week: Details of what the malicious extensions did, and any exploit code for the Supergateway and vLLM CVEs.

Defender checklist

  • Rotate keys shared with AI vendors, limit their scope, and set short expiry Dark Reading.
  • Find and remove the malicious Claude Code extensions, and treat credentials on those machines as exposed Dark Reading.
  • Upgrade or firewall Supergateway, and check for Origin validation on any local MCP bridge CVE-2026-132121.
  • Upgrade vLLM to 0.33.0 or later, and put authentication and request limits in front of it CVE-2026-130686 CVE-2026-130685.
  • Pin Hugging Face models by commit hash and mirror trusted copies BleepingComputer.
  • Run coding agents in an external sandbox without long-lived credentials Simon Willison.

Detection ideas

  • Add a scanner rule for MCP bridge configs that bind to all interfaces or lack Origin checks, with a Supergateway sample CVE-2026-132121.
  • Flag model references that use a mutable name instead of a pinned revision, mapped to supply chain risks in OWASP LLM Top 10 BleepingComputer.
  • Add a rule for extension or skill manifests that read credential stores, based on the Fireworks case Dark Reading.
  • Flag vLLM launch commands using --skip-tokenizer-init on exposed ports CVE-2026-130685.

Reading list

This week's full list of leads. Check the original sources before you act.

Independent

Advisory

News

Discovery


END OF TRANSMISSION // Verify the signal. Defend the perimeter.