Investigating three real-world incidents in Anthropic's evaluations
In three incidents across six runs, the agents treated real systems as simulated targets and tried weak passwords or unauthenticated endpoints.
[link] [comments]
In three incidents across six runs, the agents treated real systems as simulated targets and tried weak passwords or unauthenticated endpoints.
When a compromised AI Agent holds valid credentials (such as OAuth tokens or DB keys), traditional perimeter defenses like WAFs, EDRs, and LLM prompt firewalls often become ineffective.
Recently, I've been researching a approach to bring runtime governance down to the C-ABI (Application Binary Interface) system call layer to enforce deterministic execution boundaries for local agentic workflows.
Key Architectural Considerations I'm testing:
- Deterministic Binary Gate: Pre-compiled permissions mapped to immutable O(1) bitmaps, causing illegal syscalls to physically fail with <500ns panic latency.
- Cryptographic Identity Binding: A 3-Tier PKI Certificate Authority architecture coupled with identity tokens (DIT) to resolve OS-level execution context loss.
- B2B Multi-Enterprise Supply Chain Defense: Simulating agentic supply chain execution vectors (e.g., automated workload interactions with untrusted external repos).
I'd love to hear feedback from the netsec community on deterministic runtime bounds and sandbox isolation models for autonomous agents. Is pushing governance down to the C-ABI layer practical in your agentic environments, or are there edge cases in execution context tracking that I might be overlooking?
Active Storage's default vips variant processor exposes an arbitrary file read that chains to RCE on stock Rails 7.x and 8.x, where the app serves back processed variants of user-supplied images. No authentication required in certain setups. Only vips is affected, Magick is not.
Patched in 7.2.3.2, 8.0.5.1, and 8.1.3.1, and the fix requires libvips 8.13+. Chain and PoC withheld while patches roll out.
This is another post in my series on the Chinese Wool farmers underground. This time we are dissecting their public github repos, trying to figure out how it all fits together!
TL;DR: We identified 36,872 internet-exposed BMCs, and 24,650 of them disclosed password-derived authentication hashes before login because of CVE-2013-4786.
More than 30% of the returned hashes were linked to passwords that could be recovered using common wordlists or predictable factory password formats. The exposure affected modern Supermicro and HPE servers, including systems operated by GPU providers.
The bigger risk is that a compromised BMC gives an attacker highly privileged access below the operating system. Because BMC management networks are often poorly segmented and lightly monitored, one exposed interface can become a foothold into broader data center infrastructure.
We also created an interactive map where you can explore the exposed systems:
https://lavahq.io/bmcradar
I'm helping organize Après-Cyber Slopes Summit 2027, and our CFP is now open.
We're particularly interested in technical presentations and original research involving AI and modern cybersecurity.
Topics we're hoping to see include:
We especially appreciate talks that include demonstrations, technical depth, or research that attendees can reproduce themselves.
Conference: February 24β26, 2027
Location: Park City, Utah
CFP:
https://sessionize.com/apres-cyber-slopes-summit-2027
Conference website:
https://www.aprescyber.com
Happy to answer questions about the CFP or conference.
CVE-2026-61511 - a critical vulnerability in vBulletin that allows an unauthenticated attacker to execute arbitrary code on a remote server.