On Monday 20 March 2023, OpenAI took ChatGPT offline. Users had begun reporting that chat titles belonging to strangers were appearing in their own sidebars — a small, deeply unsettling failure, because the thing being leaked was the record of what people had chosen to ask a machine in private.
The post-mortem, published on 24 March, is worth reading carefully because of what it was not. There was no intrusion. The cause was a bug in redis-py, the open-source Redis client library OpenAI used: a change deployed that morning caused a spike in request cancellations, which could corrupt a pooled connection so that one user's request was answered with another user's cached data. OpenAI said the bug exposed titles from other active users' chat histories and, where two users were active simultaneously, the first message of a newly created conversation — but not the full text of existing conversations. It also disclosed that payment-related data of roughly 1.2% of ChatGPT Plus subscribers active during a nine-hour window may have been visible to another user: names, email and payment addresses, card expiry dates and last four digits. Full card numbers were never exposed.
What made March 2023 historic was not the bug. It was the response. On 31 March, Italy's data protection authority, the Garante, ordered an immediate temporary limitation on OpenAI's processing of Italian users' data — citing the 20 March breach alongside concerns about lawful basis and age verification. ChatGPT went dark in Italy for roughly four weeks. It was the first time a national regulator had pulled a consumer generative-AI product off the market on privacy grounds, and it arrived barely four months after the product launched.
For the security profession, the lesson was that AI systems concentrate an unusually intimate class of data — the questions people are too embarrassed, worried or curious to ask anyone else — inside infrastructure built at startup speed. A connection-pooling bug that would be a minor incident in most products becomes, in a chat assistant, a stranger reading the title of your private consultation.
Cl0p's GoAnywhere harvest comes due
March was when the victims of Cl0p's February campaign started surfacing. The gang had exploited CVE-2023-0669 in Fortra's GoAnywhere managed file-transfer product — disclosed 1 February, patched 7 February — and claimed on its leak site to have stolen data from more than 130 organisations, a figure never independently verified. Rubrik confirmed on 14 March that attackers had reached a non-production test environment; Hitachi Energy confirmed employee data exposure on 27 March. Others named in the window included the City of Toronto, Procter & Gamble, Saks Fifth Avenue, Rio Tinto and the UK's Pension Protection Fund. It was a complete dress rehearsal — same gang, same tactic, same product category — for the MOVEit campaign Cl0p would launch ten weeks later.
3CX surfaces on the way out
On 29 March, CrowdStrike disclosed that 3CX's Windows desktop softphone had been trojanised and was shipping a backdoor; Volexity published its own analysis the next day. Customers had been reporting endpoint-security detections since around 22 March and had largely been reassured they were false positives — a week of alerts dismissed. With 3CX marketing to roughly 600,000 customer organisations, it was immediately one of the highest-reach supply-chain incidents since SolarWinds. Its full strangeness would only emerge in April, when investigators found the attack that caused it.
Sun Pharma confirms ransomware — and a revenue hit
Sun Pharmaceutical Industries, India's largest drugmaker by revenue, disclosed an "information security incident" in a stock exchange filing on 2 March 2023, saying it had isolated the affected network and that core systems and operations were not impacted. On 24 March the ALPHV/BlackCat operation added the company to its leak site, claiming more than 17 TB of stolen data including personal files on over 1,500 US employees — claims Sun Pharma has never confirmed. Two days later, in a follow-up filing, the company confirmed the incident was a ransomware attack, said it had received a ransom demand, and warned that the attack would reduce revenue in the affected businesses, flagging further possible costs including higher cyber insurance premiums and litigation.
The gap between "core systems unaffected" on 2 March and a revenue warning on 26 March was never publicly reconciled — and that gap is precisely why the incident matters in this archive. It made Sun Pharma an early test case for how much a listed Indian company must actually say about a ransomware event, under a CERT-In regime that requires rapid reporting to the government but says little about what shareholders and the public are owed. Two years later, the DPDP framework would begin answering the second question.
⏳ Time capsule — March 2023
- Silicon Valley Bank failed on 10 March after a bank run — the largest US bank failure since 2008; Signature Bank followed on the 12th.
- OpenAI released GPT-4 on 14 March, its first flagship model able to accept image input.
- UBS agreed on 19 March to acquire Credit Suisse in a Swiss government-brokered rescue.
- "Everything Everywhere All at Once" won seven Oscars on 12 March, and Michelle Yeoh became the first Asian actress to win Best Actress.
The month AI met the regulator
Italy's four-week ChatGPT ban was the opening move in a regulatory relationship that produced the EU AI Act and, in India, the DPDP framework's eventual reckoning with automated processing. It established that AI products would be governed by existing privacy law rather than waiting for bespoke rules — a precedent every subsequent enforcement action has built on. The redis-py incident itself became the standard teaching example that AI systems fail in ordinary software ways with extraordinary privacy consequences, a theme this archive picks up again in 2026 when AI toolchains and agents become attack surfaces in their own right. And Cl0p's GoAnywhere harvest, viewed from here, reads as the trailer for the film that opened in May.