Open
Asked by Thane Ruthenis
- Current best answer
- Nobody knows. Thane Ruthenis, who asked, was unsure: it "seems possible" there will be a spiral, "or maybe not". This site's maintainer guesses the chance of a months-long, hard-to-reverse breakdown of online finance at about 1–3% over 18 months (a guess made in late October 2026). That is one person's judgment, not a forecast. The reasoning: past attacks on financial firms caused outages of days to weeks followed by recovery, and banks keep offline backups of account records. Against that, nothing like today's tools has been seen before, so the past is a weak guide. On 14 November 2026 Thane Ruthenis added that the real question is how quickly basic services could be restored offline, and that preparing for total breakdown "doesn't really seem threat-appropriate"; he also asked whether he was underrating how fragile these systems are.
Partly answered
Asked by Elizabeth
- Current best answer
- Most guides plan for three to seven days; the Red Cross figure for water and food when staying at home is about two weeks. In the best-known hack of a single financial institution, a California credit union in 2024, online banking was out for about two weeks. One commenter, Philipp Risius, suggests planning for up to two weeks without mains electricity. Another, Thane Ruthenis, doubts that preparing for a total breakdown fits the threat. Nobody has answered how much to keep for neighbors.
Partly answered
Asked by Elizabeth
- Current best answer
- Probably rarely, in the view of most of the commenters who have discussed it; none says they have tried one in an emergency. The useful part is the offline reference library it searches, which is free on its own. One commenter, lilkim2025, expects a plain search tool over the same manuals to be faster and more reliable than a small chatbot. Another, Thane Ruthenis, thinks it is not unhelpful but doesn't fit the threat well. Most of what you'd need is specific to your situation and has to be prepared in advance.
Partly answered
Asked by Afterimage
- Current best answer
- Not automatically, and people who follow this disagree. Finding a flaw first isn't enough. Defenders found the PHP2Shell flaw on August 1, yet attackers were using it by October 17, four days before a fix was released. The writer Zvi called that "an update against defenders being advantaged". Grendel1209, who described working in cybersecurity at a threat-intelligence company, expects freely downloadable AI models to favor attackers. Vitalik Buterin predicts defense wins in the end, "but the transition will be hard". CrowdStrike, a security firm, reported that in the first half of 2026, for flaws with public demonstration code, 88% of the attacks it observed came within 48 hours of that code's release.
Partly answered
Asked by johnswentworth
- Current best answer
- Two reports in September 2026 give part of the answer. CAISI, the US government's AI testing center, said on 17 September that GLM-5.3, a Chinese model anyone can download, was the most capable downloadable model at hacking so far. It also said GLM-5.3 was "significantly lower" than the best US models, about four months behind on its tests. Anthropic, an AI company, reported on 29 September that in its tests GLM-5.3 built working attacks about as often as its own model Claude Mythos Preview, announced about five months earlier: 50 of 410 attempts against 56 on one test. Anthropic also found GLM-5.3's safety limits easy to get around. So on these tests downloadable models have reached about where the leading models were a few months before, and they still trail the newest US models, some of which are released only to vetted users. Tests are not real-world attack ability. Earlier guesses in the discussion were slower. In August 2026 hadad put 85% on a downloadable model hacking as well as Mythos, a leading model, within 24 months, and Vladimir_Nesov guessed "maybe 2 years, in 2028". In October 2026 the writer Zvi judged the best downloadable model "more than six months behind". Forecasters at the forecasting site Metaculus put about 95% on one matching Mythos on a hacking test before July 2027.