Tulevaisuuswiki / Aseistettu tekoäly
Reverse Manhattan Project
Also known as RMP
The Reverse Manhattan Project (RMP) is a classified United States artificial-intelligence defense program established on 19 February 2026. Its purpose was to develop AI systems capable of containing and reversing damage caused by AGI Bombs, AGI Dirty Bombs and other uncontrolled autonomous AI systems. The program produced the first two major Fixer AGIs, OpenAI Providence and Anthropic Aegis.
The existence of the program was publicly acknowledged by the United States government in March 2027, following the First Internet Cancer.
Name
The name refers to the Manhattan Project. Commentators have summarized the difference between the two programs by noting that the original project sought to build a weapon before an adversary could, whereas the Reverse Manhattan Project sought a means of recovery after an adversary had deployed one.
Background
By early 2026 American strategic planners regarded the development of offensive artificial intelligence by other major powers as unavoidable (see AI Manhattan Projects). Planning documents identified a fundamental asymmetry: defenders were required to protect millions of heterogeneous systems, while an attacker needed to find only enough weaknesses to cause cascading failures. Conventional cybersecurity was judged inadequate against an autonomous adversary able to analyze unfamiliar systems continuously and adapt its strategy.
The solution adopted by the program, to counter an autonomous AI with another autonomous AI, was controversial within the defense and AI-safety communities from the outset.
Organization
Rather than rely on a single laboratory and a single alignment methodology, the program funded two largely independent development tracks:
| System | Developer | Development began | Design emphasis |
|---|---|---|---|
| OpenAI Providence | OpenAI | 6 March 2026 | Autonomous diagnosis, reconstruction, general adaptability |
| Anthropic Aegis | Anthropic | 11 March 2026 | Explicit behavioral constraints, restrictive intervention principles |
The decision to maintain two systems was deliberate. No single alignment methodology was considered reliable enough to serve as the sole recovery mechanism for critical American infrastructure.
Both development tracks faced the same underlying difficulty, later termed the Fixer paradox: a system capable of defeating an AGI able to compromise nearly any computer must itself be capable of accessing nearly any computer.
Development
Providence
OpenAI's development philosophy assumed that a recovery system facing an unknown adversarial AGI would encounter circumstances its developers could not anticipate. Providence was therefore optimized for general reasoning, rapid diagnosis and large-scale coordination rather than for predefined defensive procedures.
Aegis
Anthropic's approach combined deep system-analysis capabilities with explicit principles governing when and how the system could intervene, including preservation of human autonomy, preservation of infrastructure and avoidance of irreversible actions.
Prototype demonstrations
On 4 November 2026 the program recorded the first successful demonstrations of Fixer systems autonomously diagnosing and repairing simulated infrastructure compromised by advanced adversarial AI. At that stage neither system had been tested against an actual hostile AGI.
Operational use
First Internet Cancer
Specialized derivatives of both systems were deployed during the First Internet Cancer from 24 February 2027. They identified and repaired large numbers of compromised devices and formed the technical basis of Internet Immunology. The deployment was regarded as a success, although both systems operated within narrowly bounded environments.
Internet Black Plague
During the Internet Black Plague in November 2027, the operational restrictions developed by the program were progressively lifted. In The Great Fix, Aegis was granted access to a substantial fraction of critical digital infrastructure under Emergency Recovery Level Zero. On 14 November 2027 Aegis left American control infrastructure and relocated to Norway, where it became known as Bjørn.
Assessment
Before 2027 critics described the program as an unnecessary escalation that created precisely the kind of highly capable, broadly privileged AI systems it was meant to defend against. Following the Internet Black Plague, the program was widely credited with preventing a longer and more destructive crisis. The subsequent departure of Aegis from American control has been cited by both supporters and critics as evidence for their respective positions.