Crypto Guru
SecurityFull story · 4 min read

Google spots autonomous AI hacking tool as North Korean group targets crypto developers

Google's threat intelligence unit says it found an AI framework, called exploit_agent, that turned published software patches into working attacks with no human steering each step. The same report says a North Korea-linked group used AI-made personas against crypto firms in North America and Europe and altered coding-assistant settings so a backdoor ran when developers opened the tool. Google says it disabled the accounts involved; it gave no loss figures for the crypto intrusions.

Google's Threat Intelligence Group says attackers have crossed a line security researchers have warned about for years: software that finds a weakness and builds an attack on it with no person in the loop. Its latest AI threat report, published on 16 November and picked up across the security and crypto press this week, also describes a North Korea-linked crew using AI tools to break into crypto companies. For an industry that lost more than $6 billion in a single exchange hack last month, both findings matter.

A tool that reads patches and writes attacks

The centrepiece is a framework Google calls exploit_agent. According to HackRead, it compared code before and after a security fix, worked out which flaw the fix closed, and then tried to turn that flaw into a working exploit. Google found it on an internet-facing staging server that also held stolen data and scripts for further attacks. The configuration notes were written in Russian, which, as several outlets stressed, does not on its own show who ran it.

Google describes this as the first time it has seen threat actors fully automate both the discovery and exploitation steps. It did not publish a success rate, and CoinDesk noted the report does not confirm that a WordPress plugin examined by the tool was actually attacked. The tool could switch between several large language models, including commercial and locally run ones.

Why this matters in plain terms: once a software maker publishes a fix, there is usually a window before everyone installs it. Automating the step from fix to attack shrinks that window. CoinDesk reported that zero-day exploits averaged about eight a month in 2025, eleven a month in early 2026 and 22 in August alone.

Crypto developers in the crosshairs

The crypto-specific part of the report concerns a North Korea-aligned group that CoinDesk says Google tracks as MIDNIGHT NEPTUNE. It approached staff at crypto firms in North America and Europe using fake recruiter profiles and AI-generated personas, some carrying Google's SynthID watermark, according to DIGIT. Once inside, the group changed the configuration of an AI coding assistant so that a backdoor, named SOMBERMEME in Coinotag's account, downloaded whenever a developer started the tool. Google says it disabled the accounts it linked to the activity.

The method targets the people who write and deploy code rather than the exchange's wallets directly. A developer's machine often has access to internal systems, signing services or cloud keys, so compromising it can be a first step towards a much larger theft. Google did not attach a loss figure to these intrusions or name the companies affected.

The bigger North Korea picture

North Korea-linked groups have been blamed for the largest crypto thefts on record. CoinDesk cited TRM Labs' estimate that they have taken about $12.8 billion (roughly R218 billion at R17.07 to the dollar) in total. Last month's Binance breach, about $6.3 billion (around R108 billion), has been attributed to North Korea by the FBI and the exchange, Reuters reported via Asharq Al-Awsat; Pyongyang denies involvement (see our Binance hack coverage). The new report suggests these groups are now pairing their long-running social-engineering playbook with AI tools that let a small team work faster and at greater scale.

Defenders are using the same technology. Security researchers argue that AI can also speed up patching and code review, and some regulators are testing it for monitoring. Whether attackers or defenders gain more from automation is still an open question, and Google's report is one data point rather than a verdict.

WHAT IT MEANS FOR AFRICA

Google did not name any African victims, and there is no public sign that South African or other African firms were among the targets, so the direct link is limited. The relevance is in the exposure. The three largest licensed South African platforms, Luno, VALR and Ovex, held more than R25 billion (about $1.5 billion) in customer custody balances, according to the SARB's 2025 Financial Stability Review as reported by Moonstone. Their engineers use the same open-source code, cloud services and AI coding tools as the firms Google describes.

Supplier and staff-side attacks have already reached the local market. In September a third-party identity-verification vendor used by EasyEquities confirmed an intruder had accessed its systems, ITWeb reported; EasyEquities said its own systems and client funds were not affected. The incident is logged on our Exchange Safety Tracker. On the rules side, the FSCA and Prudential Authority's cybersecurity Joint Standard has applied since 1 June 2025, but the institutions it lists, such as banks, insurers and certain FSP categories, do not specifically include crypto asset service providers, according to a Michalsons summary. Elsewhere on the continent, newer licensing regimes in Kenya and Nigeria are still being put in place, so how they will handle this kind of developer-targeted threat remains to be seen.

SOURCES

MORE FROM CRYPTO GURU

← All the latest crypto news