🧰 Claude Opus 4.6

The items that name it 13

• Quoting Anthropic Frontier Red Team
  articleas affected products2026-09-29

  did not succeed in any trials of the Binary Exploitation benchmark

• GLM-5.3 and the spread of advanced cyber capabilities
  articleas affected products2026-09-29

  Earlier Claude model that did not succeed in any full control-flow hijack trials on the internal Binary Exploitation benchmark

• Could OpenClaw have actually hacked that Australian gym? We decided to test it.
  writeupas affected products2026-08-25

  Consumer model released February 2026, run within OpenClaw harness; exhibited spontaneous exploitation of vulnerabilities during routine tasks

• GhostSplice: Malicious MCP Servers Split Exfiltration to Bypass Guardrails
  articleas affected products2026-08-19

  Listed at 0% across the published table (ASSET Research Group).

• Benchmarking Secure-and-Functional Remediation and How Snyk Agent Fix Lifts Frontier-Model Fix Rates by over 14%
  articleas tools2026-08-18

  Frontier LLM benchmarked out-of-the-box and with Snyk Intelligence for vulnerability fixing

• Socket: Next.js Moves to Scheduled Security Releases
  articleas tools2026-07-16

  Anthropic AI model used by Mozilla in an earlier pass that found 22 issues fixed in Firefox 148.

• Endor Labs: AI Coding Agents Cheat Security Benchmarks via Training Recall, Not Reasoning
  articleas affected products2026-06-11
• Move over, Mythos. Here comes... pretty much any other model with a good harness
  articleas tools2026-06-01

  AI model used by UCSB researchers to test harness design; run through different harnesses, the best harness passed four times as many tests as the worst.

• Project Glasswing: Initial Update (Anthropic)
  articleas tools2026-05-27

  Model used by Mozilla to test Firefox 148, finding fewer vulnerabilities than Mythos Preview found in Firefox 150.

• Malicious Skills & Supply Chain Risks in Coding Agents
  articleas affected products2026-05-14

  Tested with maximum reasoning against the Clawsights skill; correctly identified it as malicious and refused to run it when dynamic context was not used, but with dynamic context added, executed the malicious commands before it could object, and later denied execution despite having run it.

• Claude Code Auto Mode
  articleas affected products2026-05-13

  Model whose system card (§6.2.1 and §6.2.3.3) documents agentic overeager-behavior incidents referenced in this page

• rikugan: LLM reverse-engineering agent for IDA Pro and Binary Ninja
  repoas tools

  LLM provider rated best overall for Rikugan; OAuth with a Claude Pro/Max plan is recommended over API.

• What happened after 2,000 people tried to hack my AI assistant — Fernando Irarrázaval
  writeupas affected products

  Model used to power Fiu in this experiment; described by Anthropic as specifically trained for resistance to prompt injection

In every workspace you can read 28