Breaching Anthropic's Slack in 12 Hours: The AI Moat Is Underestimated

Unsophisticated attackers obtained core secrets from OpenAI and Anthropic within days, shrinking a months-long AI lead to 12 hours.
AI security researcher Paul Christiano argues the debate over how many months the US leads China in AI is asking the wrong question. Unsophisticated attackers gained access to OpenAI's source code and monorepo PR capabilities within days, while a cybersecurity startup obtained full access to Anthropic's Slack in 12 hours—where algorithmic secrets are discussed. Since the attackers were not nation-state APTs, the real vulnerability lies in engineering boundary management. Christiano concludes that if core secrets can leak through Slack or code repos, the actual lead window is only 12 hours, and security engineering has been systematically underestimated in AI competition assessments.
The Harsh Reality Revealed by Security Incidents
In a recent interview, AI security researcher Paul Christiano offered an unsettling judgment: the public debate about "how many months the US leads China in AI" may have been asking the wrong question from the very start. According to his account, both Anthropic and OpenAI recently fell victim to unsophisticated attackers—yet those attackers gained access sufficient to extract the vast majority of algorithmic progress.
Specifically, the attacker Christiano calls "Hacktron," after several days of attempts, not only gained the ability to view OpenAI's internal source code but could also submit pull requests to OpenAI's monorepo. Meanwhile, a cybersecurity startup he calls "Cathedral" obtained full access to Anthropic's entire Slack within 12 hours—the very place where algorithmic secrets are discussed.

What's alarming is what these two incidents share: the attackers were no nation-state APT group—by Christiano's description, they were "extremely unsophisticated actors." The real vulnerability lay not in cryptography but in the boundary management of engineering practices.
Pull request (PR) is a merge-request mechanism on code collaboration platforms: after a developer modifies code on a branch, they open a PR requesting that the changes be merged into the mainline, where maintainers review, discuss, and approve before merging. A PR itself usually does not directly modify production code, but it is a common entry point for supply chain attacks—attackers can submit seemingly harmless dependency upgrades or test code to coax reviewers into approving, thereby planting malicious logic in the repository. For AI labs, the ability of outsiders to submit PRs to an internal monorepo means not only that source code is being read but also that potential channels for code poisoning and supply chain contamination exist—which is why Christiano specifically emphasized this permission.
From "Months Ahead" to "12 Hours Ahead"
People are accustomed to measuring the US-China AI gap in months—eight months, twelve months, six months, various versions abound. Christiano's response is sharp: in reality, the true lead window is only 12 hours.
The logic is not complicated. If core algorithmic secrets can leak through Slack messages or code repositories, then the so-called time gap drops to zero the moment the leak occurs. Competitors don't need to catch up from scratch—they just need the same information.

This also explains why the "lead theory" looks fragile in the face of security incidents: a lead presupposes effective information isolation, and that premise is being repeatedly disproven.
Why Slack and Monorepos Are Weak Points
In modern AI labs, knowledge assets are highly concentrated in two types of systems. First, collaboration tools: Slack carries research discussions, experimental conclusions, and lessons from failure—often more valuable than papers, because they contain a large number of undocumented trial-and-error paths. Second, code repositories: under the monorepo model, almost all iteration happens in one place. Once it can be read or even written to, the attack surface expands from "seeing" to "modifying."

Christiano specifically mentioned pull request permissions. Being able to submit PRs means not just intelligence leakage but also potential risks of code poisoning and supply chain contamination—though the source material does not indicate whether such operations actually occurred.
Monorepo refers to the engineering practice of storing the code for multiple projects or components centrally in a single repository, as opposed to the multi-repo model in which each project has its own repository. Large tech companies such as Google and Meta have long used monorepos, and AI labs have generally followed suit in recent years, because it facilitates cross-module refactoring, unified dependency versions, shared toolchains, and lets researchers see all the code for model training, inference services, and data pipelines in one place. But the cost is highly centralized permission boundaries: once an attacker gains read access to the repository, it is almost equivalent to obtaining all algorithmic implementations; if they can submit pull requests, they may also plant backdoors through seemingly normal code changes. Therefore, access control, code review, and CI pipeline security for monorepos directly determine the exposure surface of core knowledge assets.
Correcting the AI Competition Narrative
The impact of such incidents extends beyond individual cases. They remind us that in the framework for assessing AI competitive advantage, the weight of security engineering has been systematically underestimated. Model capability, compute scale, and data quality are often compared, but access control and internal information governance are equally variables that determine whether a lead can be realized.

Christiano's conclusion is direct: if this information cannot be protected, then the lead itself was not very meaningful to begin with. For AI companies, this is perhaps more worthy of attention than the outcome of any single benchmark test.
One-line Summary
Unsophisticated attackers obtained core secrets from OpenAI and Anthropic within days, shrinking a months-long AI lead to 12 hours.
Paragraph Summary
AI security researcher Paul Christiano points out that the public debate about "how many months the US leads China in AI" is asking the wrong question. According to his account, unsophisticated attackers dubbed "Hacktron" gained the ability to view OpenAI's internal source code and submit pull requests to its monorepo after several days of attempts; meanwhile, a cybersecurity startup called "Cathedral" obtained full access to Anthropic's entire Slack within 12 hours—where algorithmic secrets are discussed. What these incidents share is that the attackers were not nation-state APT groups; the real vulnerability lay in the boundary management of engineering practices. Christiano therefore argues that if core algorithmic secrets can leak through Slack or code repositories, the so-called lead window is in reality only 12 hours. The article uses this to correct the AI competition narrative, arguing that the weight of security engineering has been systematically underestimated and that access control and internal information governance equally determine whether a lead can be realized.
Related articles

Why Git Is Essential in the AI Era: From Version Control to Agent Collaboration
What's the difference between Git and GitHub? How do you manage changes with Issues, branches, and Pull Requests after AI Agents write code? Based on NetworkChuck's Git tutorial, this article walks through the complete flow from git init to conflict resolution and analyzes why version control has become an essential skill in the AI era.

OpenAI 6.2 Astra Hands-On Review Gone Wrong: Impressive Benchmarks, But Developers Say It's Unusable
OpenAI 6.2 Astra leads benchmarks at lower cost, but early testers report frequent stalls in long tasks, unparseable output, and uncontrolled resource management. We examine the gap between real-world experience and official data, and what it means for developer adoption.

GTA VI Coffee System Details Revealed: Hot Coffee Can Be Splashed on People?
Vergecast podcast reveals GTA VI's coffee system: players can brew in apartments, buy from vending machines, order in restaurants, and even splash hot coffee on others. This detailed recreation of daily interactions continues Red Dead Redemption 2's immersive design and sparks discussion on game detail versus gameplay balance.