On 20 November 2026, Amazon Threat Intelligence, led by CISO CJ Moses, reported that a China-linked operator used Cairn, an open-source autonomous AI penetration-testing framework, against universities in East Asia. The Record's coverage says Amazon began investigating on 10 September, after finding a publicly reachable server hosting a Cairn instance with its prompt history exposed. Canada was not named as a victim. The relevance is a policy one: Ottawa's June 2026 strategy promotes open-source AI and, by design, carries no binding AI-specific rules.
What Amazon actually found
The operator wrote prompts in Simplified Chinese, and timestamps matched China Standard Time. The operator supplied IP addresses and told Cairn to "obtain shell access." Per Moses, the agents then planned, executed and adjusted "without the operator choosing each next step."
Most attempts failed. One run took 148 actions over 16 hours and failed, including an attempted Log4Shell exploit. Another spent more than six hours testing over 30,000 credential combinations. One run did succeed. The agents found and exploited a path traversal flaw in an open-source medical imaging AI application. Amazon assesses with high confidence that this was a zero-day, because no advisory or fix existed. The same flaw was then used against a second university.
The application is MONAILabel. Its maintainers published advisory GHSA-qc22-rmgc-vg94 on 18 November 2026, rating it critical. It affects version 0.8.5 and earlier, and the advisory notes that authentication is disabled by default. As of our reading, the advisory lists no patched release yet; a fix is merged in a pull request and awaits the next version.
According to The Record, Google has reported Cairn use against US aerospace firms, and Gambit Security reported a financially motivated group using it against online retailers. So this is no longer a single-actor anecdote.
The strongest case for restricting open release
The case for regulators is serious. Cairn is general-purpose, so defenders cannot predict or fingerprint its behaviour easily. Agents that run for 16 hours straight can outpace a security operations centre's triage capacity, as Moses warned. If a tool lowers the skill needed to find zero-days, a government could reasonably want licensing, know-your-customer rules or liability for publishers. Canada's strategy promotes open-source development without binding obligations, so a critic could call it naive.
Why that remedy fails on the evidence
The incident points elsewhere. Look at where the attack succeeded. It did not succeed through some novel model capability that open release unlocked. It succeeded against an open-source application that shipped with authentication off by default and an unvalidated file path. Most of Cairn's other attempts failed, even over long runs. The decisive variable was a defender's unpatched, exposed software.
The operator also made a basic mistake. The Cairn server was publicly reachable, and that is how Amazon saw the full prompt history. Openness helped the defenders here. Because the framework is inspectable, researchers can study how it behaves, and Amazon's visibility turned a covert campaign into a documented one.
Restricting publication would also be poorly targeted. Autonomous offensive frameworks are not unique to open source. Google grouped Cairn with other tools in this class, according to The Record. A state-linked operator can build or buy what it needs. A Canadian rule would burden domestic researchers and red teams, who use these tools for authorised testing, while barely inconveniencing the actor in this story.
What Canada's strategy gets right, and what it lacks
The federal government announced AI for All on 4 June 2026. Its news release says the strategy makes "security and safety paramount" and commits to modernising privacy and online safety law. Miller Thomson's analysis says the strategy does not introduce comprehensive AI legislation, and that the earlier Artificial Intelligence and Data Act is unlikely to be revived. The same analysis notes that Pillar 6 promotes open-source AI development and that Pillar 1 aims to protect critical systems from cyberattacks.
That posture is defensible. Comprehensive AI statutes risk fixing the wrong target. The Cairn case shows that the risk is a cybersecurity problem, and existing tools address it better than a new AI-specific regime.
But the strategy has a gap. It treats open-source promotion and cyber resilience as separate pillars. The Cairn campaign sits where they meet, in the open-source projects that universities, hospitals and research labs deploy. A proportionate response would be targeted:
- Secure-by-default funding. If Ottawa funds open-source AI, it should attach conditions for authentication on by default, dependency audits and a disclosure process for the projects it supports.
- Faster vulnerability handling. The Canadian Centre for Cyber Security and research institutions should have a clear path to report and track flaws in widely used open-source AI tooling.
- Defender-side AI. Pillar 1's cyber research should fund automated detection and patching. A threat that runs for 16 hours needs a response that is not purely human.
- Sector baselines. Hospital and university IT, where imaging tools sit, need minimum configuration standards, which are cheaper than regulating a code repository.
The bottom line
The Amazon report is evidence about speed and exposure, not a case against openness. It shows that autonomous agents can chain reconnaissance, credential attacks and exploit discovery with little human direction, and that the weak point was software left open by default. Canada can keep its pro-open-source stance and its light legislative touch, but only if it backs the stance with deployment security. Without that, the next Cairn-style campaign will find the same kind of target, and the policy argument for heavy regulation will get easier to make.