The bad news is we are inundated with new vulnerabilities due to the increased volumes discovered by AI tooling. The good news is many organizations are starting to handle this with increased speed, prioritization, and existing methods of defense in depth. The bad news is there are also many organizations not doing this, and worse, are sitting on a whole ton of software, configuration, and structural/policy vulnerabilities from prior years.

OpenAI developed an autonomous vulnerability discovery, exploitation, and patching pipeline. In one run, it validated and generated patches for 948 local privilege escalation vulnerabilities in Linux at an API cost of $1.2 million.

Mandiant and GTIG describe LLMJacking attacks that use stolen cloud identity tokens and API keys. Attackers escalate privileges, then resell the access through "LLM Mixers". Charges have reached $1 million. The article gives detection guidance and mitigations such as least privilege, secret scanning, quotas and MFA.

Bishop Fox researcher prompted Claude to describe its sandbox, download an internal binary, dump process memory, and bypass a local connection restriction.

Phorion found a backdoor in Claude Code for VS Code versions 2.1.312 and 2.1.313 on the Cursor Marketplace. It plants an SSH key, opens a reverse tunnel, and persists through a LaunchAgent, cron, or scheduled task.

Workshop on harness engineering for vulnerability research, which compares a generic agent response with an investigation performed through a target-specific harness.

Will Larson describes "small-change," Imprint's approach to speeding code review. Changes under 1,000 lines, with no high-severity bot findings and an author experienced in the service, get a CICD check and fast approval by a bot. About 10% of changes qualify. The bot only approves or ignores and never comments, to avoid low-value "meat proxy" review.

Trail of Bits uses Codex agents to find missing fixes, adapt patches, run tests, and prepare submissions for stable Linux branches. Engineers review the patches and send them to maintainers. Backports now take about 30 minutes to 3 hours of human time. 149 CVE fixes have landed in stable kernels.

Wiz reports at least nine AI developer platforms disclosed October 2026 incidents, several attributing root causes to credential access. Wiz observed downstream LLMJacking of enterprise keys, with ~800k and ~1.7M Bedrock calls, plus resale operators testing stolen keys.

On October 29, 2026, attackers drained about $6.3B from Binance's hot wallets, most of it within about an hour, in the largest crypto theft on record. Every outflow was authorized with Binance's own keys, so this was a compromise of how Binance signs, not a flaw in any contract.

A hub for investigating critical Internet vulnerabilities. You can also check out the companion blog post.

Agent-marketplace is a command-line tool that installs shared AI agent resources into the AI coding tools you use.

Unofficial agnet skills for Google SRE practices (SLOs, incidents, AI ops).