Anthropic started red teaming new Mythos models, first results

Anthropic began red teaming its new Mythos model, codenamed Claude Oceanus-v1-p, on June 3, with early access granted to security testers. The preview candidate focuses on advanced reasoning, coding, and cybersecurity, building on April's Mythos Preview, and a wider rollout is expected in the second half of June. Early unconfirmed reports indicate Oceanus outputs significantly surpass current models, though Anthropic has cautioned against premature celebration, citing the need for verifiable safety methods.

Anthropic appears poised to advance its most closely watched frontier work toward release. A model identifier, claude-oceanus-v1-p https://x.com/synthwavedd/status/2062519972379652339?ref=testingcatalog.com , appeared on June 3 in the Claude Console, with red teamers reportedly granted access around the same day. Oceanus seems to be the next step in the Mythos line, building on April's Mythos Preview, with a focus on advanced reasoning, coding, cybersecurity, and long-horizon agentic work rather than chat. The "-v1-p" tag indicates a preview candidate moving through evaluation, not a research artifact. Red-team access has typically preceded a wider rollout by a week or two, suggesting a plausible launch in the second half of June, close to when OpenAI's rumored GPT-5.6 codename kindle-alpha is also expected. Early discussions, including creative venues like VoxelBench, suggest that Oceanus outputs are significantly superior to those of current models, even with minimal effort, though nothing has been independently confirmed. The leap looks promising, though not yet settled. Earlier reports initially linked Mythos to Claude Code and Claude Security https://www.testingcatalog.com/anthropic-prepares-mythos-1-for-claude-code-and-claude-security/ , targeting developers and security teams. Whether it will also reach business, personal, or Max tiers, or remain enterprise-only, is unclear; Pro almost certainly waits. Join Dev Mode Discord for more Join https://discord.gg/DevMode?ref=testingcatalog.com This development coincides with Anthropic's new Institute paper, which argues that AI is already accelerating AI development, citing Mythos Preview achieving a 52x training-optimization speedup. Notably, the company frames this as a cautionary note, emphasizing that the field has not yet achieved recursive self-improvement and calling for verifiable methods to slow down, rather than celebrating prematurely.