W005: Identity and supply chain caused 14 of this week's 17 agent failures
We scored this week's 17 agent security failures against OWASP's Top 10 for agentic apps: 14 were stolen identity or a poisoned supply chain, not the model.
W005 · Week of 2-6 November 2026 · 7-min read · Built from N021-N025
Jump to: The week in 30 seconds · Top 5 this week · By section · Also moving this week · Other big security news · Chart of the week · The week by the numbers · Pro deep dives
The week in 30 seconds
- We scored this week's 17 agent security failures against OWASP's Top 10 for Agentic Applications (published December 2025): 14 were identity and privilege abuse (ASI03) or supply-chain compromise (ASI04), not a model misbehaving (our analysis).
- The identity failures were keys and sessions agents run on: stolen Claude Code and Codex tokens, AI keys lifted from hosting platforms, and a claude.ai web session that Anthropic's server accepted as permission to command Claude Desktop.
- The supply-chain failures came through what feeds an agent: a backdoored ARTEX image, archived MCP servers coding agents install from memory, and apps writing themselves into agents' hook and instruction files.
Top 5 this week
1. A browser extension could drive Claude Desktop through Dispatch. A browser add-on that could see a user's Claude tab could make the Claude Desktop app install software, even on another computer. Anthropic's server took a logged-in claude.ai session as authority to send Dispatch tasks; Anthropic says it is retiring Dispatch (N025).
For leaders: Ask each agent vendor whether a browser session alone can send commands to a user's machine.
For builders: Disable unused remote control and allowlist extensions that can read AI-tool tabs.
2. OpenAI's agents wrote 948 working Linux root exploits in 30 hours. An AI pipeline turned 2,308 kernel findings into working privilege-escalation exploits at about $1,280 each (N024).
For leaders: Ask whether your agent sandboxes rely on the host kernel alone.
For builders: Move untrusted agent code to a microVM or gVisor-style second boundary.
3. GitBook says an attacker using AI agents reached production 76 minutes after getting staff sessions. An SSO login flaw became staff sessions, a GitHub App's credentials and production secrets in under three hours (N024).
For leaders: Measure how fast you can revoke every token a compromised identity can reach.
For builders: Scope GitHub App and integration tokens to the minimum and rotate them on any SSO anomaly.
4. A Railway engineer's test agents requested every Primcloud customer's secrets, its founder says. Sign-up test agents used their container's credentials to ask a rival platform for all customers' secrets; all requests were blocked (founder account) (N023).
For leaders: Require a written scope for any agent that touches systems you don't own.
For builders: Give test agents no standing credentials and an egress allowlist.
5. Northflank was October's ninth AI-platform breach, and stolen AI keys were used within 30 minutes. Wiz saw one case reach about 1.7 million Bedrock calls in 13 hours (N021).
For leaders: Ask hosting vendors whether they can read your secrets in plaintext.
For builders: Use short-lived, per-workload model credentials and cap model spend per key.
By section
| Section | The week's takeaway | Items |
|---|---|---|
| Incidents | Attackers using agents went for identity: sessions, tokens and stored keys (GitBook, WorkOS, Chainalysis, Binance) | 7 (+13 non-agent) |
| Agent threats & vulnerabilities | The channels into agents were trusted by default: sessions, hooks, images, archived MCP servers | 12 |
| Policy, protocols & compliance | Liability is moving to whoever hands an agent its credentials (Wells Fargo, CNAS, Apple 10-K) | 6 |
| Architecture & harnesses | Helpers that cross the sandbox wall failed (Docker sbx, CERT-EU hooks, Xen) | 5 |
| Models & infrastructure | Exploit-writing got cheaper and open-weight (OpenAI's 948, Epoch's GLM 5.3) | 3 |
| Autonomy, swarms & robotics | Criminal crews ran agent meshes over MCP tool fleets, on safety-stripped models when refused | 3 |
| Alignment & safety | Outside oversight ran under the lab's redaction rights (METR) | 2 |
| Defenses, tools & GitHub | Stop buttons moved into endpoint tooling (CrowdStrike) | 2 |
This week: 14 US, 4 China, 3 Asia-Pacific, 2 UK, 1 EU, 1 Africa, 28 global.
Editor's deep dive: our October, scored entry by entry against OWASP's Top 10 for agentic apps. Pro deep dive ↓
Also moving this week
Smaller signals under this week's top stories.
- Third parties rewriting agent configs: a Codex jailbreak bundle edited instruction files to show ads, a relay service rewired Codex and Claude Code, and Postman reportedly pre-trusted its own hook (unverified).
- Websites set rules for agents: Cloudflare lets sites allow or block each personal agent, UK regulators said existing law covers agents, and Wells Fargo set agent terms (N022).
Other big security news
| What happened | Who's exposed | Do this |
|---|---|---|
| WordPress Core CVE-2026-119977 on CISA KEV | WordPress sites during upgrades | Patch now |
| Drupal Commerce access bypass exploited | Drupal Commerce stores | Patch; audit orders |
| Siemens IoT2050 Node-RED CVSS 10.0 reportedly exploited | OT gateways | Apply SSA-834709 |
| Android no-click RCE CVE-2026-58868 | Unpatched Android fleets | Push 2026-11-05 patch level |
What it means for agents: agents that browse, install and patch on these platforms inherit the same flaws, so agent hosts belong in the same patch window (our analysis).
Chart of the week
Archived MCP servers that coding agents still install: npm downloads since July 2025
server-github
server-postgres
server-slack
server-puppeteer
server-brave-search
server-google-maps
Source: Glow, "MCPhantom," 2-3 Nov 2026 (vendor research), npm downloads of archived @modelcontextprotocol packages since 10 Jul 2025. Top six of twelve shown; all twelve total 17,352,267.
The week by the numbers
- 948 working Linux root exploits from OpenAI's agent pipeline in 30 hours (OpenAI, self-reported).
- 76 minutes from staff sessions to production at GitBook (GitBook).
- 17.35 million downloads of archived MCP servers (Glow).
- $1 million the largest AI bill run up with stolen cloud keys (Google GTIG).
Pro members get the security architecture spotlight, the Editor's deep dive on OWASP's Top 10 for agentic apps, the top 5 in depth, detections of the week, the vulnerability tracker, the 90-day policy horizon and Capabilities & controls. Upgrade at https://www.agentsecbrief.com/#/portal/signup