Saturday, 21 Nov 2026 Breaking: Virginia reports first measles-associated child death of 2026, US fatalities rise to 12
BREAKING: Apple launches AI platform | Tesla earnings beat estimates | Nvidia stock surges | Bitcoin crosses major resistance
Business

OpenAI strengthens safeguards against AI model distillation, reports sharp drop in reasoning extraction attacks

OpenAI strengthens safeguards against AI model distillation, reports sharp drop in reasoning extraction attacks

New Delhi [India], November 21 (ANI): OpenAI has strengthened safeguards in its latest AI models to prevent unauthorised attempts to extract their internal reasoning and replicate advanced capabilities, reporting a sharp decline in the success rate of such attacks following security upgrades.
In an update outlining its efforts to combat adversarial distillation, the company said the average success rate of attacks attempting to extract protected reasoning from an earlier request fell to 0.2 per cent in GPT-6.2 Astra, compared with 54 per cent in GPT-6 Astra under its original safeguards.
The success rate for attacks targeting reasoning generated in the current request also declined to 0.9 per cent from 22.9 per cent, according to internal adaptive evaluations involving automated red-teaming.
Distillation is a machine-learning technique through which one AI model learns from data generated by another. While the technique has legitimate applications, unauthorised distillation can allow another model to reproduce advanced capabilities without inheriting the safety protections of the original system.
OpenAI warned that this could create risks in sensitive areas such as cybersecurity and biological research, where capabilities developed for defensive or legitimate purposes could be repurposed for harmful activities.
The company said it had placed particular emphasis on protecting its models' chain-of-thought, or the intermediate reasoning steps generated while working towards an answer. Such information can contain details not included in the final response and can provide valuable training material for competing models.
An internal study found that AI models trained on exact reasoning traces generally performed better in cybersecurity benchmarks after supervised fine-tuning than those trained on reasoning summaries. The advantage persisted at higher inference budgets following additional reinforcement learning.
To address the threat, OpenAI is combining model-level safeguards, input and output classifiers, API restrictions and investigations into coordinated extraction campaigns.
The company said it had detected multiple unauthorised distillation campaigns on its platform this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts. The pace of such activity increased significantly over the summer.
OpenAI also said it shares indicators of compromise with other AI model providers to help identify and disrupt similar activities across platforms.
Highlighting the broader national security implications, the company said foreign actors could reproduce advanced capabilities developed by US AI companies and deploy them without comparable safeguards.
It called for stronger coordination among AI developers, deployment partners and governments, arguing that protecting frontier models from unauthorised capability extraction would be essential as advanced AI systems become more widely available. (ANI)

Tags

Related News

Oracle says OpenAI contract fully funded, targets over USD 30 billion free cash flow in FY29
Business
Oracle says OpenAI contract fully funded, targets over USD 30 billion free cash flow in FY29

<p>New Delhi [India], November 21 (ANI): Oracle has secured the funding required to fulfil its infrastructure contract w...

Frontier breached, Anthropic mathematician claims solution of Riemann hypothesis with help from Claude AI
Business
Frontier breached, Anthropic mathematician claims solution of Riemann hypothesis with help from Claude AI

<p>Santa Clara (California) [US], November 21 (ANI): One of the last frontiers of Mathematics appears to have been breac...

Anthropic shares fall nearly 9% to USD 88.65 a day after Nasdaq debut
Business
Anthropic shares fall nearly 9% to USD 88.65 a day after Nasdaq debut

<p>New York [US], November 20 (ANI): Shares of Anthropic fell 8.77 per cent to USD 88.65 on Friday, a day after the make...

Jewellery trade body alerts BIS over suspected gold adulteration, urges tighter checks
Business
Jewellery trade body alerts BIS over suspected gold adulteration, urges tighter checks

<p>New Delhi [India], November 20 (ANI): The All India Gem and Jewellery Domestic Council (GJC) said it has alerted the...

Nirmala Sitharaman chairs ninth pre-Budget consultation with housing, urbanisation and culture stakeholders
Business
Nirmala Sitharaman chairs ninth pre-Budget consultation with housing, urbanisation and culture stakeholders

<p>New Delhi [India], November 20 (ANI): Union Finance and Corporate Affairs Minister Nirmala Sitharaman on Friday chair...

Cyber frauds biggest challenge for banking sector, institutions must share security strategies: DFS Secretary
Business
Cyber frauds biggest challenge for banking sector, institutions must share security strategies: DFS Secretary

<p>Mumbai (Maharashtra) [India], November 20 (ANI): Cybersecurity and cyber frauds have emerged as the biggest challenge...