The launch was a small piece of theatre that went wrong. WikiLeaks had promised a live press briefing for the morning of 7 March 2017, then announced that its stream was under attack and simply published: 8,761 documents and files it called Year Zero, the opening instalment of a series named Vault 7, drawn — it said — from an isolated, high-security network inside the CIA's Center for Cyber Intelligence at Langley. The decryption passphrase it tweeted was a line attributed to John F. Kennedy, about splintering the CIA into a thousand pieces and scattering it to the winds. The agency would not confirm that any of it was genuine; American officials would eventually describe it as the largest unauthorised disclosure of classified information in the CIA's history.
The files read like the internal wiki of a mid-sized software company — version histories, testing complaints, developer banter — except that the products were implants for iPhones, Android handsets, routers and televisions. Notes from a 2014 workshop with Britain's MI5 described Weeping Angel, named for the Doctor Who monsters that move only when nobody is watching: installed on a Samsung F8000 television, it held the set in a "Fake-Off" mode that looked like standby while the microphone recorded the room. The engineers' own list of problems still to fix recorded the giveaway — a blue LED on the back that stayed lit. Early headlines said encryption was broken; the documents showed something narrower and worse. An agency that owned the phone did not need to break Signal or WhatsApp at all.
The CIA confirmed nothing, saying only that the public should be "deeply troubled" by disclosures designed to damage the intelligence community's ability to protect Americans. Apple said many of the iOS flaws in the inventory were already fixed in the current release. On 9 March Julian Assange offered technology companies early access to the technical details; several, wary of handling stolen classified material, kept to their ordinary reporting channels. The month kept circling the same argument. On 8 March, at a Boston College conference, the FBI's James Comey said there was "no such thing as absolute privacy in America"; on 15 March the Justice Department charged two Russian FSB officers, and two hackers hired by them, over the 2014 theft of data on 500 million Yahoo accounts — the first criminal cyber charges against serving Russian officials.
The drops settled into a weekly rhythm. Dark Matter, on 23 March, described implants living in the firmware of Macs and iPhones, beneath the reach of a reinstall; Marble, on 31 March, published 676 source files of an obfuscation framework whose test strings included Chinese, Russian, Korean, Arabic and Farsi — proof of false-flag capability, WikiLeaks suggested, though researchers read it as ordinary tradecraft for hiding an author, not framing one. The hunt had meanwhile turned inward almost at once: by 13 March federal agents held a warrant for the Manhattan apartment of Joshua Schulte, a software engineer who had left the Center for Cyber Intelligence the previous November. He denied being the source for as long as the question was open. A jury convicted him in July 2022; in February 2024 he was sentenced to 40 years.
The patch nobody took a bow for
Microsoft had done something unprecedented in February 2017: it cancelled the month's Patch Tuesday outright, citing a last-minute issue, and left administrators guessing. When the updates returned on 14 March, the batch included security bulletin MS17-010, rated critical — a fix for flaws in SMBv1, the elderly Windows file-sharing protocol, across every supported version of the operating system. The bulletin was unusual in one small way: it credited no researcher for the discovery. A month later, on 14 April, the Shadow Brokers published a working exploit for exactly those flaws — EternalBlue, from the National Security Agency's own stolen toolkit — and The Washington Post later reported that the NSA, knowing what was gone, had warned Microsoft itself. Fifty-nine days after the patch shipped, on 12 May, WannaCry used EternalBlue to encrypt its way through every network that had not applied it — the subject of this archive's restoration of three days in May. March is the month the warning was issued, in writing, to everyone.
The header that opened Equifax
The Apache Software Foundation fixed a flaw in Struts, a framework holding up thousands of corporate web applications, on 6 March 2017; the vulnerability — CVE-2017-5638, a crafted Content-Type header the framework would obligingly execute — was publicly detailed on 7 March. Exploitation began almost immediately: Cisco's Talos researchers reported live attacks within about forty-eight hours, and on 8 March US-CERT circulated a notice urging organisations to apply the fix. At the credit bureau Equifax that notice arrived on schedule and went round internally on 9 March; the account later assembled for congressional investigators records that the company's scans failed to find the vulnerable system — and that on 10 March someone outside was already testing the hole in its online dispute portal. The looting came later: between 13 May and 30 July intruders extracted personal records that Equifax, announcing the breach on 7 September, first put at 143 million people and finally at 147.9 million. The patch had existed the whole time. It took two months for the failure to become theft, and six for anyone to be told.
2.2 million records on an open shelf
The trouble with the McDelivery app was not that somebody had broken in; it was that the door had never been locked. Fallible, a small Indian security startup, found that an API endpoint serving McDonald's India's ordering service would return a customer's record — name, email address, phone number, home address, home coordinates, links to social profiles — to anyone who asked, without any authentication, and that customer IDs were plain sequential integers. Walk the numbers and the roughly 2.2 million registered users fall out one by one. Fallible wrote to the company on 4 February 2017 and received an acknowledgement from a senior IT manager on 13 February. Then, for a month, nothing visible happened. On 18 March the researchers published, and the story ran across the Indian press the next day.
McDonald's India's public response emphasised what the app did not hold — no card details, no wallet passwords, no bank account information — and offered the assurance that the service "has always been safe to use", which answered a question nobody had asked. Whether anyone hostile had walked the sequence before Fallible did is unknowable, and that is precisely the point: this was an exposure reported by researchers, not a confirmed theft, and no log was produced either way. The company said the flaw was fixed; Fallible responded that the first fix was incomplete and the endpoint was still leaking. In the India of March 2017 no law required McDonald's to tell its 2.2 million customers anything at all — no breach-notification clock, no data-protection statute. Both would come years later, and this archive records their arrival.
Kaggle, Distill and 1,440 cores
At Cloud Next in San Francisco on 8 March, Fei-Fei Li — newly installed as Google Cloud's chief AI scientist — announced that Google was buying Kaggle, home to the world's largest community of data scientists, and unveiled a Video Intelligence API that could label what happens in a video, frame by frame. The talent market moved the same month: Andrew Ng announced he was leaving the chief scientist's post at Baidu, where he had built one of the field's most watched labs. Distill launched too, a peer-reviewed journal backed by Google Brain, OpenAI and DeepMind, built for the web on the wager that machine learning could be explained as carefully as it was published. And on 24 March OpenAI showed that evolution strategies — an old, almost naive optimisation method — could rival reinforcement learning once spread across 1,440 processor cores, teaching a simulated humanoid to walk in about ten minutes; read from 2026, that preference for scale over cleverness looks like the coming decade's thesis, stated early.
The defender as the attack vehicle
On 22 March the Israeli firm Cybellum disclosed DoubleAgent: register a rogue verifier DLL under Microsoft's Application Verifier, a legitimate debugging tool shipped with every version of Windows since XP, and the operating system will inject it into a chosen process at every launch, surviving reboots and reinstalls. The demonstration target was anti-virus software itself, Norton included — the guard recruited as the intruder's persistence mechanism. Malwarebytes and AVG had patched by disclosure and Trend Micro's fix was in the works; Symantec countered that no Norton vulnerability was involved and that the technique already required administrator rights, at which point an attacker scarcely needed it. Cybellum's sharper observation was that Microsoft's protected-processes design blocked the trick — and that almost no vendor beyond Windows Defender had adopted it. The next day Google's Ryan Sleevi proposed that Chrome progressively distrust Symantec-issued TLS certificates, citing an investigation that had grown from 127 questioned certificates to at least 30,000; Symantec strongly objected and called the move irresponsible. By late 2018 Chrome trusted none of them, and the certificate business belonged to DigiCert.
⏳ Time capsule — March 2017
- The Nintendo Switch went on sale worldwide on 3 March — a home console that travelled — launching alongside The Legend of Zelda: Breath of the Wild.
- On 2 March the price of one bitcoin passed the price of an ounce of gold for the first time; on 10 March the SEC rejected the Winklevoss brothers' proposed bitcoin ETF, and the price fell hard before resuming its year-long climb.
- Theresa May's letter invoking Article 50 was delivered to European Council president Donald Tusk on 29 March, starting the two-year clock on Britain's exit from the European Union.
- On 30 March SpaceX launched the SES-10 satellite on a Falcon 9 first stage that had already flown a year earlier — the first re-flight of an orbital-class rocket booster — and landed it at sea again.
What the warnings were worth
The leak's endings took seven years to arrive. Joshua Schulte, convicted in July 2022, was sentenced in February 2024 to 40 years. The CIA's own WikiLeaks Task Force report — written in 2017, released in redacted form in June 2020 — called the security of the mission network "woefully lax" and conceded the agency could not say what had been taken: its estimate ran from 180 gigabytes to 34 terabytes, because the system kept no user-activity logs. Julian Assange pleaded guilty in June 2024 and flew home to Australia. And the television that listened stopped being a scandal and became a product category; the argument Weeping Angel started is now had about every speaker, doorbell and assistant sold.
The two warnings were honoured mostly in the breach. Fifty-nine days after MS17-010 came WannaCry; six weeks after that, NotPetya rode the same exploit through networks that still had not patched. Equifax told the world in September about intruders whose first visit its own investigators dated to 10 March, and settled with American regulators in 2019 for up to $700 million. The pattern March 2017 first drew — a fix published, a gap unclosed, a component nobody remembered installing — became the industry's standing nightmare, rehearsed again at scale when Log4j broke in December 2021. India, in time, got the laws the McDelivery episode showed it lacked: a six-hour reporting clock in 2022, a data-protection act in 2023. The warnings still arrive on schedule; the editions that follow record what they were worth.