Insights & Research
Vulnerability research, AI security, and field notes on compliance from the team that breaks systems so attackers can't.
CVE-2026-59822: MCP Authentication Bypass in LiteLLM
LiteLLM's MCP endpoint caught a rejected credential and continued with an anonymous session. The strict check still ran and still failed correctly. Its failure just stopped being terminal.
GHSA-2xp9-vwfh-vxw4: AVIF Decoding RCE in Next.js Image Optimization
Next.js shipped two critical RCEs on 25 August. The one with the CVE number needs Windows and a specific router setup. The one without it needs an AVIF file and a check that was written as a TODO comment in 2018.
Disclosure Used to Buy You Days. Now It Buys You Hours.
An OCaml maintainer watched automated probes hit his server ten minutes after opening a bug fix, then had an AI agent build a working exploit in under a minute. Two CVEs from this year show the same pattern already running against production software.
CVE-2026-1678: DNS Parser Overflow in Zephyr
A stale bounds value in Zephyr's DNS name parser lets a remote attacker overflow a 255-byte buffer with five 63-byte labels. The check is there. It just doesn't update as the buffer grows.
GHSA-c677-q3wr-gggq: Remote DoS in Valkey's Cluster Bus
A malformed PING packet crashes a Valkey cluster node by exploiting an out-of-bounds read in clusterIsValidPacket. One packet, no authentication, full node crash.
GHSA-hjr9-wj7v-7hv8: Unauthenticated DoS in Sliver C2's HTTP Listener
Three bugs chain together to crash Sliver's HTTP listener without authentication. A breakdown of the vulnerable code, the encoder bypass, and why a single HTTP request can exhaust server memory.