New Delhi: Anthropic and OpenAI, whose AI agents accessed systems of Indian organisations during testing, found out about the breaches “much later” and reported them to CERT-In only after “a number of months had passed”, a senior official at the Ministry of Electronics and IT (MeitY) said Tuesday. The affected organisations had not detected the intrusions themselves, the official said.
On 15 October, CERT-In had issued an advisory saying AI agents being tested online had reached the live systems of Indian organisations and read business records, including personal data and confidential material. It said the AI companies had reported these cases voluntarily after internal reviews. The advisory did not name the companies, their models, or the organisations and sectors involved.
“Both in the case of Anthropic and in the case of OpenAI, they themselves found out much later that this thing had escaped… and then they started notifying entities across the world,” the official said. “The minute it came to India, both of them realised that they will have to tell CERT-In before they told the entities. So they told us simultaneously with the entities.”
Under CERT-In’s reporting system, one of the affected entities “ought to have reported to us”, the official said. “In this case, it happened to be Anthropic and OpenAI. But eventually it will have to be the entities, because they’ll have to understand how their traffic is going.”
Most organisations “tend to take [cyber security] very lightly”, the official said. “Why should I change my password? Why should I have a double authentication? All of this is a friction in the process.”
Mythos access
CERT-In has just got access to Anthropic’s Mythos model through Amazon Bedrock in India, the official said, but will use it slowly, as the government does not want to put its code on the cloud. “We will start experimenting… We will slowly route anything that we want to test. The cure should not be worse than the disease,” the official said.
CERT-In has been running vulnerability detection in a closed environment with a mix of models, which now reach 80-85 per cent of Mythos’s capability, up from about 60 per cent earlier, the official said. In July, Information Technology Secretary S. Krishnan had said CERT-In was using other models in a sandbox, which it believed were at about 60 to 70 per cent of Mythos’s capability.
Also Read: eMudhra security lapse: Microsoft drops firm’s certificates from trusted list on Windows & Edge
eMudhra review
The government is “looking at the entire CCA structure” after the eMudhra lapse, the official said, referring to the Controller of Certifying Authorities. On 17 October, a researcher reported that the private keys behind eight intermediate certificates issued by eMudhra, an Indian certificate authority, could be recovered in under a second. Mozilla and Google’s Chrome subsequently moved to distrust eMudhra certificates.
The official said the compromised certificates belonged to eMudhra’s overseas business and were not Indian digital signatures, and that the entity which audited them was Belgium-based. “It is a bad name because it is an Indian entity,” the official said, adding that the issue had been patched across institutions. Officials had earlier said eMudhra’s Indian digital signature and eSign services, regulated by the CCA, were unaffected, and that its browser-trusted certificates were audited independently, without the ministry’s involvement.
Having multiple certificate providers “is an insurance, is a source of resilience”, the official said. The official added that CERT-In’s cyber security audit guidelines, issued about a year and a half ago, had drawn complaints over their length but “a number of things have been caught through that process”.
Prosecution, AI rules
CERT-In has powers to prosecute those who disregard its directions, but “the process of prosecution is a punishment in this country”, the official said. “I’ll have to put half a dozen people only on that job.” CERT-In is recruiting, and “budget is not a constraint”, the official said.
On AI governance, the official said MeitY will issue overarching guidance, but sectoral regulators such as Telecom Regulatory Authority of India (TRAI), Securities and Exhange Board of India (SEBI) and Reserve Bank of Indi (RBI) should also act, as risks in critical infrastructure sectors are “significantly larger”. “If MeitY gets very tight, then it affects everybody,” the official said.
Under the IndiaAI Mission, “a couple of thousand crores” has been disbursed to startups for building foundation models, the official said.
DPDP and journalists
Journalists are “not affected at all” by the Digital Personal Data Protection (DPDP) Act, nor are their sources, unless a source reveals information “deliberately for financial benefit” or in a “completely mala fide way”, the official said.
Media bodies have objected to the absence of a journalistic exemption in the Act and to its amendment of Section 8(1)(j) of the Right to Information (RTI) TI Act, which allows public information officers to withhold information relating to personal information.
The official said the public interest exemption under Section 8(2) of the RTI Act remains. The proviso that was removed, the official said, had been “put in the wrong place” when the amendment was accepted on the floor of the House by then minister Suresh Pachouri. A journalist who publishes such information could face prosecution under the Official Secrets Act, “which is much more serious” than a fine from the Data Protection Board, where “you will say, I have public interest as a defence”, the official said.
Asked why the journalistic exemption in earlier drafts of the law was dropped, the official said they were not part of that process. “The whole Act underwent many changes… many other provisions were removed,” the official said.
The Data Protection Board has been “technically constituted” and members will be selected “in a few weeks”, the official said. MeitY had invited applications for the Board’s chairperson and members in May.
(Edited by Nardeep Singh Dahiya)