The Security Field for Enterprise AI. Governing access, protecting agent actions and validating the systems behind enterprise AI.
Joined September 2026
- Tweets25
- Following10
- Followers9
- Likes11
🔍 The way in wasn't the model.
Testing in @AnthropicAI's researcher program, Cathedral says its AI agents got into some systems in under 12 hours via third-party business software.
Exclusive: Cathedral, a months-old cybersecurity start-up, recently told Anthropic that its researchers had penetrated sensitive internal data and messaging systems at the AI company, according to a draft report reviewed by The Post. wapo.st/48fN3us
2/ @AnthropicAI says they reached limited sales, HR and IT systems, not Claude chats, code or weights.
They also says Cathedral didn't disclose all its testing, as program rules require. Cathedral says it kept data access minimal.
A researcher's lab test: a hacked device got full control on a target server via Tailscale SSH, where rules let it in with forwarding.
No patch or CVE yet. @Tailscale's engineering VP says it's fixing issues, but is "not recommending that everyone disable Tailscale SSH."
‼️ WARNING - A researcher just dropped a report on Tailscale SSH flaws — one compromised device could get root on others, and there’s NO PATCH yet.
In lab tests, SSH port forwarding enabled privilege escalation and lateral movement across already-permitted devices.
Tailscale says a fix is coming.
Read - thehackernews.com/2026/10/un…
👉 @Tailscale's advice for now: turn off SSH forwarding if you don't need it
(set TS_SSH_DISABLE_FORWARDING=1, then restart tailscaled), and tighten SSH rules.
Tailscale SSH is turned on per device. The report says forwarding can't be turned off on headscale 0.29.5.
LiteLLM published how its v1.107.0 release was compromised. If your team runs AI agents, it's worth a read.
According to LiteLLM, it started with a master key that was accidentally pushed to a public repo in May. Nobody turned it off.
We have completed our investigation into the attack chain behind the v1.107.0 release. Read it here: docs.litellm.ai/blog/securit…
2/ The key unlocked an internal AI agent platform. That platform held a GitHub token with admin power over LiteLLM's whole org, and it handed the token over.
Releases usually needed approval from an account tied to an office device. That week, any 2 engineers could approve one.
AI agents under evaluation reached live systems in India and read business records, including personal data, per CERT-In.
Finding a credential is not authorization to use it.
We read the advisory. Here's what it means for agent security.
🧵:
11/12
@AgentGuard_AI also supports component review before trust & audit evidence during operation.
Review skills, plugins & MCP components for documented risks. Trace runtime decisions and approvals.
A component scan and a runtime control address different parts of the risk.
12/12
Deploy AgentGuard on the agents you operate, alongside network isolation, scoped identities and secret rotation.
Start with one workflow. Verify which actions are observed, approved or blocked before expanding.
Learn more: AgentGuard.AxoraSecurity.com…
