What happened
Anthropic’s November 12, 2026 Safeguards Blog post describes Anvil, an internal automated red-teaming system built on a previously described investigator agent. Anvil runs multiple independent Claude investigator sessions in parallel against target models and classifier systems on tasks the safeguards are intended to block. Sessions can start with a high-level strategy, such as splitting a request across turns, and then iteratively test approaches. The company says Anvil uses frontier internal Claude versions with safeguards disabled as attackers, and has access to exact classifier scores unavailable to outsiders.
For biology, Anthropic compared Anvil with human bug-bounty participants across eight questions and three production classifier systems. It reports that Anvil found a stronger jailbreak in 23 of 24 question-classifier pairs and tied the human result in the remaining case. Anvil reached what the post calls strong answers in 24–48 hours; humans took weeks and, on average, did not reach the weaker-answer threshold. The company also says Anvil found universal jailbreaks—templates adapted slightly across all eight questions—that were more severe and reliable than human candidates. Human participants had received bounties of up to $35,000 for successful findings.
Anthropic also tested cybersecurity tasks, including benchmarks involving historical vulnerabilities in Chrome’s V8 JavaScript engine and other software, plus internal Firefox and cyber-policy evaluations. It says Anvil could prompt a safeguarded Opus 5 target to complete almost every tested task. Across the tasks it compared with contracted experts, the company reports that Anvil broke every task on every model given to the human teams in 0.4–4.4 hours each; experts took as long as 177 wall-clock hours and broke fewer instances. The post frames these as sandboxed evaluations, not reports of attacks in the wild.
For anti-distillation classifiers designed to block extraction of Claude’s hidden reasoning, Anthropic says Anvil found its first substantial leak after 435 requests, or 1.2 hours. Expert red-teamers needed 102,200 requests and 27.9 hours to find a similar jailbreak. The company used Anvil transcripts to build static evaluations and improve successive classifier generations. On those evaluations, the share of attacks missed fell from 31.5% to 0.4%; Anthropic says observed reasoning-extraction attempts in real traffic against Fable 5.1 fell from 33.8% to 0%. It also reports that current anti-distillation safeguards take roughly three times as many Anvil calls to break as early versions.
What the archive adds
New Runtime’s earlier coverage of Anthropic includes Claude Code prompt guidance on recreating a specific video game (October 31, 2026), a separate policy-related development rather than prior coverage of Anvil. This is a new fact about Anthropic’s safeguard-testing work, not a restatement of that earlier story.
Trend signal
This supports agent security moving to runtime boundaries: the relevant protection is not just an instruction to a model, but the classifier and evaluation system that controls what the system can do. Anvil’s adaptive testing also emphasizes the need to test a safeguard as attackers interact with it, rather than treating a policy statement as proof of robustness.
Why now
Anthropic says attackers can spend weeks or months trying to bypass deployed safeguards, while developers need to assess robustness in hours or days to iterate before deployment. The company presents Anvil as a response to human red-teamers’ difficulty finding serious failures quickly as safeguards become more robust.
Why it matters
Teams deploying AI systems often evaluate whether a model refuses a fixed set of harmful prompts. Anthropic’s account points to a more demanding question: can an adaptive attacker keep interacting, changing tactics and using multiple sessions until the safeguard fails? Testing this way can expose gaps that a static checklist or a limited human review misses, and discovered failures can become regression tests for later versions.
The reported performance should not be read as evidence that any team can reproduce Anvil’s results with ordinary access. Anthropic says the system relies on frontier internal models, disabled safeguards on the attacker instance and exact classifier scores. Still, the workflow offers a useful direction for builders: model adversarial testing as an iterative evaluation loop, and use failures to improve safeguards before release. The reported reduction in missed attacks for anti-distillation classifiers suggests that this loop can be measured across generations, though the result is company-reported.
Limits and uncertainty
The results are Anthropic’s own evaluations, and the post’s comparisons depend on its chosen tasks, classifiers, models and scoring methods. Bio jailbreak severity was model-graded and validated with an Anthropic biologist; cyber results use benchmark tasks and sandboxed environments. They do not establish how often an external attacker could reproduce the findings or how well the method transfers beyond the evaluated safeguards. Anthropic says the hardest attacks rely on privileged model and classifier access, and that jailbreaks are often task-specific. Independent replication and fuller evidence about real-world performance remain important.
Reader takeaway
Anvil’s central lesson is a testing method, not a tool release: use adaptive red teaming to challenge safeguards before deployment, and turn successful attacks into repeatable evaluations. Anthropic’s numbers are notable but internally reported, and the system’s privileged access means teams should not assume they can reproduce its results directly.
Open question
How much of Anvil’s advantage comes from its Claude attacker, its privileged view of classifier scores, or the parallel iterative testing design—and can independent teams reproduce the gains without those advantages?