Anthropic Allows Glasswing Partners to Share Mythos-Based Findings
When Anthropic early last month unveiled Mythos, the high-flying AI company limited access to the frontier model for fear that its advanced capabilities in detecting software vulnerabilities could be used by bad actors to quickly create exploits for software flaws it finds.
Those given access – including Microsoft, Google, JPMorganChase and Nvidia – were allowed to use Mythos to develop stronger defensive security protections against AI threats as part of Project Glasswing, an initiative put together by Anthropic. They were also barred from sharing the findings from their work.
According to reports, the AI vendor revised its position and is allowing the Glasswing members to share their findings, tools, or code they’ve developed using Mythos.
“We fully support our partners sharing findings and companies outside of Glasswing to triage vulnerabilities,” an Anthropic spokesperson told Reuters in a statement. “While there was never a specific Glasswing NDA, confidentiality protections were something partners asked for at the outset and were built into agreements partners signed.”
The spokesperson added that “as the program has matured, we’ve adapted them to ensure key information can be shared broadly – including outside the program – for maximum defensive impact.”
‘Defensive AI Cannot Scale in Isolation’
According to some cybersecurity pros, Anthropic’s decision to revise the agreements is an indication that the vendor learned that it didn’t make sense to keep the findings of the Glasswing members secret.
“Anthropic’s policy shift for Project Glasswing recognizes that defensive AI cannot scale in isolation,” said Jacob Warner, director of IT for Xcape. “When Claude Mythos Preview debuted with the ability to autonomously chain complex vulnerabilities, hoarding those findings among an elite tier of tech firms created an untenable security imbalance.”
Phil Wylie, senior consultant and evangelist at Suzu Labs, echoed the sentiment, saying that the decision to loosen the restrictions around sharing findings “reflects an important reality in modern defense: siloed threat intelligence has limited value when organizations are facing the same systemic risks.”
Will Accelerate Coordinated Defense
“Allowing Project Glasswing participants to share research, tools, and indicators with regulators, industry groups, and the broader security community could help accelerate coordinated defense and vulnerability remediation across critical sectors,” Wylie said. “What makes this notable is that it signals growing recognition that AI-assisted security research needs responsible collaboration frameworks, not just strict containment policies.”
Now comes the challenge of managing the fallout of such information becoming more widely available. According to Wylie, there needs to be a balance in transparency via safeguards to ensure that sensitive research from Glasswing can’t be weaponized before organizations have time to patch found vulnerabilities and respond to the threats they may create.
It Will Bring Challenges
XCape’s Warner noted that the policy change can lead to a deluge of highly complex AI-generated disclosures that security teams made up of humans will not be able to triage manually.
“To navigate this new landscape, security teams must move away from manual code reviews and implement automated validation pipelines,” he said. “Organizations should immediately establish structured intake channels for AI-derived disclosures, require external researchers to provide machine-readable proofs-of-concept, and upgrade staging environments to handle automated regression testing.”
They also need to prioritize the automation of internal patch validation pipelines to prevent defensive workflows from collapsing under the expected massive volumes of AI-discovered flaws, Warner added.
Reaction to Daybreak?
However, Lydia Zhang, co-founder and president of Ridge Security Technology, said that “it’s more believable that Anthropic is opening up Mythos due to pressure from the release of OpenAI Daybreak.”
OpenAI, which soon after Anthropic announced Mythos, released its own cybersecurity-focused AI model, GPT-5.4-Cyber, earlier this month, announced Daybreak, its own security effort that comprises frontier AI model capabilities and the vendor’s Codex Security – an agent used to detect, validate, and patch vulnerabilities – to enable enterprises to address security flaws before bad actors can exploit them.
“Even though [Anthropic’s] approach of sharing vulnerabilities with organizations that may face similar risks sounds reasonable and well-intentioned, in reality, how can outsiders know which organizations may face the same vulnerabilities in the first place?” Zhang asked. “Also, allowing [a Glasswing member] to act ‘at its own discretion’ creates a major gray area, as there are no clear regulations, laws, or industry conventions defining what that discretion should be based on. The bigger concern is that competition between AI models could put sensitive information at stake.”
That said, to Warner, Anthropic’s decision makes sense.
“For a project named Glasswing, keeping your vulnerability findings trapped in a locked vault was a surprisingly opaque strategy,” he said.
